<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: borodutch.eth 💎🤲</title>
    <description>The latest articles on DEV Community by borodutch.eth 💎🤲 (@backmeupplz).</description>
    <link>https://dev.to/backmeupplz</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F852251%2F54cbb5f3-9548-41dd-9d8d-c09bbbad6b1f.jpg</url>
      <title>DEV Community: borodutch.eth 💎🤲</title>
      <link>https://dev.to/backmeupplz</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/backmeupplz"/>
    <language>en</language>
    <item>
      <title>What a 1% creator fee works out to, with real numbers</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Mon, 16 Nov 2026 10:30:23 +0000</pubDate>
      <link>https://dev.to/backmeupplz/what-a-1-creator-fee-works-out-to-with-real-numbers-47a1</link>
      <guid>https://dev.to/backmeupplz/what-a-1-creator-fee-works-out-to-with-real-numbers-47a1</guid>
      <description>&lt;p&gt;Most creator platforms tell you their fee in one number and leave the rest in the footnotes. Sendclap's is one number too, 1%, so here is exactly what that comes to, what it doesn't include, and where it stops being the cheapest option.&lt;/p&gt;

&lt;h2&gt;
  
  
  The arithmetic
&lt;/h2&gt;

&lt;p&gt;Sendclap takes 1% of each payment. No monthly fee, no setup fee. What you keep depends on how the buyer paid:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Payment&lt;/th&gt;
&lt;th&gt;Sendclap 1%&lt;/th&gt;
&lt;th&gt;You keep by card&lt;/th&gt;
&lt;th&gt;You keep in crypto&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;$0.10&lt;/td&gt;
&lt;td&gt;$9.31&lt;/td&gt;
&lt;td&gt;$9.90&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$50&lt;/td&gt;
&lt;td&gt;$0.50&lt;/td&gt;
&lt;td&gt;$47.75&lt;/td&gt;
&lt;td&gt;$49.50&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;$100&lt;/td&gt;
&lt;td&gt;$1.00&lt;/td&gt;
&lt;td&gt;$95.80&lt;/td&gt;
&lt;td&gt;$99.00&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The card column is lower because cards go through Stripe, and Stripe charges its own processing fee on your own Stripe account. The table uses Stripe's US price of 2.9% + 30¢. Yours depends on your country and the card type, and the &lt;a href="https://sendclap.com/tools/stripe-fee-calculator" rel="noopener noreferrer"&gt;Stripe fee calculator&lt;/a&gt; shows US, UK and EU cards. Sendclap's 1% sits on top of that, it doesn't replace it.&lt;/p&gt;

&lt;p&gt;Crypto payments cost only the 1%. The buyer pays from any wallet, the payment is converted to USDC on Base, and a public smart contract forwards 99% to your wallet and 1% to Sendclap in one transaction.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the 1% does not cover
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Taxes.&lt;/strong&gt; Sendclap is not a merchant of record. You handle your own sales tax and VAT, the same as you would taking payments through Stripe directly. Merchants of record such as Lemon Squeezy do that work for you and charge more for it. If you don't want to deal with VAT, that is a real reason to pay more.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your Stripe rate.&lt;/strong&gt; Covered above. It is Stripe's number, not ours.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network fees for the buyer.&lt;/strong&gt; Crypto buyers pay gas on Base, which is usually small, but it is theirs.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How that compares
&lt;/h2&gt;

&lt;p&gt;We read every competitor's fee from its own pricing page and date it on &lt;a href="https://sendclap.com/compare" rel="noopener noreferrer"&gt;the comparison page&lt;/a&gt;. Short version, as of mid-November 2026: Patreon takes 10% plus processing, currency conversion and payout fees. Buy Me a Coffee takes 5% plus Stripe's processing and a payout fee. Ko-fi takes 0% on simple tips and 5% on memberships, products and other paid features, with card processing on top. Gumroad, Lemon Squeezy and Substack have their own structures, and each has &lt;a href="https://sendclap.com/compare/patreon" rel="noopener noreferrer"&gt;its own page&lt;/a&gt; with what you keep on $5, $10 and $50.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where 1% is not the lowest
&lt;/h2&gt;

&lt;p&gt;Be skeptical of anyone who says "lowest fee" without a footnote. Ko-fi's free tier, GitHub Sponsors for individuals and Liberapay take 0% on donations, and Stripe's own Payment Links cost nothing beyond Stripe. If all you need is a tip jar, one of those is cheaper.&lt;/p&gt;

&lt;p&gt;The 1% is for people who want memberships, tips and digital products with license keys in one page, with card and crypto, and who are paying 5 to 12% elsewhere for the same features. It is also non-custodial: card money goes from Stripe to you, crypto goes from the contract to your wallet, and Sendclap never holds it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try the numbers on your own prices
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://sendclap.com/tools/fee-calculator" rel="noopener noreferrer"&gt;fee calculator&lt;/a&gt; takes your price and payments per month and shows what you keep on each platform. If your numbers say another platform is the better fit, believe them.&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>payments</category>
      <category>opensource</category>
      <category>indiehackers</category>
    </item>
    <item>
      <title>How to speed up or cancel a stuck Ethereum transaction in Plain Wallet, and the 25% fee rule behind it</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Sun, 15 Nov 2026 11:00:19 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-to-speed-up-or-cancel-a-stuck-ethereum-transaction-in-plain-wallet-and-the-25-fee-rule-behind-1573</link>
      <guid>https://dev.to/backmeupplz/how-to-speed-up-or-cancel-a-stuck-ethereum-transaction-in-plain-wallet-and-the-25-fee-rule-behind-1573</guid>
      <description>&lt;p&gt;When an Ethereum transaction sits pending because its fee is too low, the fix is to send a new transaction with the same nonce and a higher fee, and Plain Wallet now does that for you: open &lt;strong&gt;Transactions&lt;/strong&gt; on the home screen, tap &lt;strong&gt;Speed up&lt;/strong&gt; to resend the same transaction, or &lt;strong&gt;Cancel transaction&lt;/strong&gt; to replace it with a transfer of nothing to yourself. Both ask you to approve exactly what gets signed, and both bump every fee field by at least 25%. That is the whole feature, and it arrived in &lt;a href="https://plainwallet.app" rel="noopener noreferrer"&gt;Plain Wallet&lt;/a&gt; 0.5.9 to 0.5.12 on 2 November 2026. The rest of this post is how it works underneath, with the numbers from the code at version 0.6.1 (commit &lt;code&gt;00b0fd6&lt;/code&gt;, read on 2026-11-15), because "just raise the gas" hides three details that decide whether the replacement is accepted at all.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why does a transaction get stuck, and what does "same nonce" mean?
&lt;/h2&gt;

&lt;p&gt;Every account has a counter, the nonce, and the network runs an account's transactions strictly in order: nonce 7 cannot be included before nonce 6. A transaction whose maximum fee per gas is below what blocks currently charge is not rejected, it waits in the pool, and everything behind it from the same account waits too. You cannot delete a waiting transaction. You can only send another one at the same nonce that the network prefers, and whichever of the two is included first counts.&lt;/p&gt;

&lt;p&gt;"Prefers" has a precise meaning. A node keeps the waiting transaction unless the new one pays enough more. Plain Wallet's own test comment puts the requirement at at least 10% on each fee field, and the node's refusal when you fall short is the string &lt;code&gt;replacement transaction underpriced&lt;/code&gt;. Plain Wallet matches that text, and the older &lt;code&gt;replacement fee too low&lt;/code&gt;, in a function called &lt;code&gt;feeRefusal&lt;/code&gt; in &lt;code&gt;lib/chain.ts&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Plain Wallet do when a node says "replacement transaction underpriced"?
&lt;/h2&gt;

&lt;p&gt;This is where it started, in 0.5.9. You send a transaction while another one from the same account still waits at that nonce, and the node refuses. Now the wallet recognises the refusal, offers the transaction again with 25% higher fees, and asks before signing. It does this both for requests from sites and on the Send screen.&lt;/p&gt;

&lt;p&gt;Version 0.5.10 widened the same idea to the other fee refusal, &lt;code&gt;max fee per gas less than block base fee&lt;/code&gt; (and &lt;code&gt;transaction underpriced&lt;/code&gt;), which means the network got more expensive between the moment you reviewed the transaction and the moment you approved it. There the retry is not a replacement, so the fee is raised to what the network charges now and never lower than what you already offered. Anything else, such as &lt;code&gt;nonce too low&lt;/code&gt; or &lt;code&gt;insufficient funds&lt;/code&gt;, is not a fee problem and is shown as it is.&lt;/p&gt;

&lt;h2&gt;
  
  
  How much higher is the new fee? The 25% rule, with numbers
&lt;/h2&gt;

&lt;p&gt;The function that decides is &lt;code&gt;raise&lt;/code&gt; in &lt;code&gt;lib/chain.ts&lt;/code&gt;. For a replacement it multiplies every fee field by 125 and divides by 100 in bigint arithmetic, then takes the larger of that and what the network charges now. For a plain retry it skips the 25% and takes the larger of your old value and the current one. I ran the actual function from the 0.6.1 source in Node 22 on 2026-11-15, with a 21,000-gas transfer that was first sent at a max fee of 20 gwei and a priority fee of 1 gwei (a total of 0.00042 ETH at most):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Situation&lt;/th&gt;
&lt;th&gt;New max fee&lt;/th&gt;
&lt;th&gt;New priority fee&lt;/th&gt;
&lt;th&gt;Most it can cost&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Replace, network now charges 18 gwei and 0.5&lt;/td&gt;
&lt;td&gt;25 gwei&lt;/td&gt;
&lt;td&gt;1.25 gwei&lt;/td&gt;
&lt;td&gt;0.000525 ETH&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Replace, network now charges 30 gwei and 2&lt;/td&gt;
&lt;td&gt;30 gwei&lt;/td&gt;
&lt;td&gt;2 gwei&lt;/td&gt;
&lt;td&gt;0.00063 ETH&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Retry after a fee refusal, network now 24 and 1.5&lt;/td&gt;
&lt;td&gt;24 gwei&lt;/td&gt;
&lt;td&gt;1.5 gwei&lt;/td&gt;
&lt;td&gt;0.000504 ETH&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Retry after a fee refusal, network now 15 and 0.5&lt;/td&gt;
&lt;td&gt;20 gwei&lt;/td&gt;
&lt;td&gt;1 gwei&lt;/td&gt;
&lt;td&gt;0.00042 ETH&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The first row is the case where the market is cheaper than your old fee and the 25% rule still applies: the node will not swap your transaction for one that pays less than 10% more, so the wallet pays 25 gwei even though 18 would clear the market. The second row is the opposite case, where fees jumped past the +25% and the wallet follows the network. The last row shows that a retry never goes below what you offered. The wallet raises the max fee, the priority fee and the legacy gas price together, so no field is left below the bump. (viem fills in &lt;code&gt;gasPrice&lt;/code&gt; next to the EIP-1559 fields, so a transaction can carry both.)&lt;/p&gt;

&lt;p&gt;The "most it can cost" column is the number you see before you approve. The fee is shown again when you open the review, not copied from the first attempt, because the wallet asks the network for current fees when you tap the button.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does Cancel work?
&lt;/h2&gt;

&lt;p&gt;Cancel is a replacement whose content is nothing. Plain Wallet takes your pending transaction, keeps its nonce and chain, changes the recipient to your own address, sets the value to zero and removes the call data, and sends that at the higher fee. If the network includes it first, your original transaction can never run, because its nonce is used up. If the original was already about to be included, it wins, and the screen says so up front: &lt;em&gt;"If it is included first, the original never happens."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;One detail I did not expect to need: the gas limit of the cancel is asked from the node with &lt;code&gt;estimateGas&lt;/code&gt; for a zero-value transfer to yourself, rather than assumed to be 21,000. Some chains, Arbitrum among them, count more than 21,000 for a plain transfer, and a hard-coded limit would make the cancel fail there.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where do the pending transactions come from?
&lt;/h2&gt;

&lt;p&gt;The wallet keeps a list of everything it broadcast, on every network, in the browser's local storage on that device and nowhere else. Each entry is the signed transaction itself plus the network name, the time and the site that asked. The sender, the nonce and the chain are read back out of the signature, so there is no second copy of those that could disagree with what is signed. The last 100 are kept, and older ones are on a block explorer.&lt;/p&gt;

&lt;p&gt;To say where each one stands, the list looks for a receipt by hash. Without one, it asks how many transactions the sender has had included. If that count is still at or below the transaction's nonce, the transaction is &lt;strong&gt;Pending&lt;/strong&gt;. If the count has moved past the nonce and there is still no receipt, something else took that nonce, and it is shown as &lt;strong&gt;Replaced or dropped&lt;/strong&gt;. Once a fate is known it is stored, so the network is not asked again. When you speed up or cancel, the old and new transactions share a nonce, and only the newest of them gets the buttons, which stops you from stacking a third replacement on a second one by accident.&lt;/p&gt;

&lt;p&gt;The buttons are hidden for watch-only accounts, because their key lives on another device and there is nothing here to sign with.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does "already known" mean?
&lt;/h2&gt;

&lt;p&gt;In 0.5.12 I fixed something that looked like a failure but was not. If you send a transaction, close the wallet, and send the identical one again, the node answers &lt;code&gt;already known&lt;/code&gt;: it already holds those exact bytes. Signing is deterministic, so the same nonce and the same fees produce the same signature. Before the fix the wallet reported an error. Now it counts the transaction as sent, lists it under Transactions with its hash, and lets you speed it up. The wallet matches three phrasings of that answer: &lt;code&gt;already known&lt;/code&gt;, &lt;code&gt;known transaction&lt;/code&gt; and &lt;code&gt;already imported&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would tell someone before they tap Speed up
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The new transaction does not stop the old one, it races it. Whichever is included first counts, so a speed-up is safe, and a cancel is a request.&lt;/li&gt;
&lt;li&gt;Check the "most it can cost" figure. A replacement is at least 25% more than the first attempt.&lt;/li&gt;
&lt;li&gt;The list of sent transactions stays on the device and is never synced. A speed-up or cancel goes to the RPC of the network you chose, like any other transaction.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The feature shipped in four small releases in an afternoon, &lt;a href="https://git.borodutch.com/backmeupplz/plainwallet/releases/tag/v0.5.9" rel="noopener noreferrer"&gt;0.5.9&lt;/a&gt; to &lt;a href="https://git.borodutch.com/backmeupplz/plainwallet/releases/tag/v0.5.12" rel="noopener noreferrer"&gt;0.5.12&lt;/a&gt;. Everything is in the open at &lt;a href="https://github.com/backmeupplz/plainwallet" rel="noopener noreferrer"&gt;github.com/backmeupplz/plainwallet&lt;/a&gt;, and the wallet itself, for Chrome, Firefox, Safari, Android, iPhone and Mac, is at &lt;a href="https://plainwallet.app" rel="noopener noreferrer"&gt;plainwallet.app&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions people ask
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Can I cancel an Ethereum transaction?&lt;/strong&gt; Only by replacing it: a zero-value transfer to yourself at the same nonce and a higher fee. If the original is included first, it still happens.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How much higher does the gas fee need to be to replace a transaction?&lt;/strong&gt; Nodes need at least 10% more on every fee field. Plain Wallet uses 25%, or the current network fee if that is higher.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does Plain Wallet send my transaction list anywhere?&lt;/strong&gt; No. It is stored only on the device; only a transaction you approve is sent, to the RPC of the selected network.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published on the &lt;a href="https://blog.plainwallet.app/speed-up-cancel-stuck-ethereum-transaction/" rel="noopener noreferrer"&gt;Plain Wallet blog&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ethereum</category>
      <category>web3</category>
      <category>cryptocurrency</category>
      <category>opensource</category>
    </item>
    <item>
      <title>How Tugport's API serves service status from memory instead of asking Docker every time</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Sat, 14 Nov 2026 10:30:18 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-tugports-api-serves-service-status-from-memory-instead-of-asking-docker-every-time-1278</link>
      <guid>https://dev.to/backmeupplz/how-tugports-api-serves-service-status-from-memory-instead-of-asking-docker-every-time-1278</guid>
      <description>&lt;p&gt;Tugport's API answers "what is running and in what state" from memory instead of asking Docker each time you ask. Since v0.1.28 (26 October 2026) the lists behind the panel, the disk usage numbers and the "is there an update" check are served from values the server already holds, and refreshed in the background. On 14 November 2026, against the live panel on v0.1.35, listing every service took the same time as the cheapest call the API has, and the difference between them was smaller than the noise.&lt;/p&gt;

&lt;p&gt;I did not set out to make the API fast. I set out to stop the panel from feeling slow, and it turned out the fix was mostly about who is allowed to wait.&lt;/p&gt;

&lt;h2&gt;
  
  
  What was slow before v0.1.28
&lt;/h2&gt;

&lt;p&gt;Three things in the panel sounded cheap and weren't:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Service state.&lt;/strong&gt; Every project and service list needed to know whether each container was running, so each request listed the containers on every server. A comment in the code puts one such listing at about 100 ms on a busy server, and a page that loads a project list and a service list pays it twice.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Docker disk usage.&lt;/strong&gt; The Servers page shows how much space images, build cache, volumes and containers take. Docker has to size every volume to answer that, which the code comment says takes seconds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The update check.&lt;/strong&gt; Asking GitHub for the newest release on every visit to Settings would burn through the 60 unauthenticated calls an hour that GitHub allows, and it adds a round trip to someone else's server.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of these values changes between two clicks of the same person. That is the whole idea behind the change in &lt;a href="https://github.com/backmeupplz/tugport/commit/b460853" rel="noopener noreferrer"&gt;commit b460853&lt;/a&gt; (the pull request is #39, ten files, 107 lines added).&lt;/p&gt;

&lt;h2&gt;
  
  
  Service state: the proxy already knows
&lt;/h2&gt;

&lt;p&gt;Tugport's proxy has to know every container to route traffic, so it rebuilds its routing table whenever a container starts, dies, stops, pauses or unpauses on any server, and every 30 seconds as a safety net. Each rebuild already lists every service's containers. The change keeps that listing, as a shared snapshot next to the proxy, and the API reads it instead of asking Docker again (&lt;code&gt;known_replicas&lt;/code&gt; in &lt;code&gt;src/runtime.rs&lt;/code&gt;). Before the proxy has built its first table, it falls back to a fresh listing, so a freshly started server never answers with nothing.&lt;/p&gt;

&lt;p&gt;One detail mattered more than it looks. The proxy used to rebuild on &lt;code&gt;start&lt;/code&gt;, &lt;code&gt;die&lt;/code&gt;, &lt;code&gt;stop&lt;/code&gt;, &lt;code&gt;pause&lt;/code&gt;, &lt;code&gt;unpause&lt;/code&gt; and health changes. Once the API reads the same snapshot, a service you just created or deleted has to show up in the lists too, so the commit adds the Docker &lt;code&gt;create&lt;/code&gt; and &lt;code&gt;destroy&lt;/code&gt; events to the ones that trigger a rebuild. Without them the panel would have shown a new service only after the next 30-second tick.&lt;/p&gt;

&lt;h2&gt;
  
  
  Disk usage and updates: serve old, refresh behind
&lt;/h2&gt;

&lt;p&gt;For the two values that are slow to compute and fine to be a few minutes old, the commit adds a small helper in &lt;code&gt;src/util.rs&lt;/code&gt;, &lt;code&gt;cached&lt;/code&gt;. It keeps a value with the time it was fetched, and when a value is older than its limit it does something that is easy to get wrong: it still returns the old value immediately, and starts one background fetch to replace it. Only the very first request for a key waits for a real fetch.&lt;/p&gt;

&lt;p&gt;Two choices in it are worth copying:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The entry is marked fresh the moment a refresh starts&lt;/strong&gt;, so ten requests arriving together start one fetch, not ten.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A failed refresh keeps the old value.&lt;/strong&gt; The entry was already marked fresh, so the next attempt happens after another full interval instead of on every request.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The limits are five minutes for Docker disk usage and ten minutes for the update check. At startup the server also measures its own Docker usage in the background (&lt;code&gt;warm_docker_usage&lt;/code&gt;), so even the first person to open the Servers page after a restart gets an instant answer. The commit adds a test, &lt;code&gt;cached_serves_stale_while_refreshing&lt;/code&gt;, that checks the sequence: first fetch waited for, fresh value served without fetching, stale value served while the fetch runs, new value once it lands.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I measured on the live panel
&lt;/h2&gt;

&lt;p&gt;I called the panel's API through Radiant's key (an operator on one project) from the worker that writes these posts, 60 rounds, one call to each endpoint per round so that network drift hits all of them equally, on 14 November 2026 at 10:02 UTC, with the panel on v0.1.35 (&lt;code&gt;GET /api/about&lt;/code&gt; reports &lt;code&gt;0.1.35&lt;/code&gt;, commit &lt;code&gt;3166248&lt;/code&gt;).&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Endpoint&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;th&gt;Fastest&lt;/th&gt;
&lt;th&gt;Median&lt;/th&gt;
&lt;th&gt;90th percentile&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;/api/about&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;returns the version&lt;/td&gt;
&lt;td&gt;61 ms&lt;/td&gt;
&lt;td&gt;97 ms&lt;/td&gt;
&lt;td&gt;118 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;/api/me&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;says who the key is&lt;/td&gt;
&lt;td&gt;54 ms&lt;/td&gt;
&lt;td&gt;87 ms&lt;/td&gt;
&lt;td&gt;103 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;/api/projects&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;projects with service states&lt;/td&gt;
&lt;td&gt;55 ms&lt;/td&gt;
&lt;td&gt;89 ms&lt;/td&gt;
&lt;td&gt;104 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;/api/services&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;every service with its state&lt;/td&gt;
&lt;td&gt;61 ms&lt;/td&gt;
&lt;td&gt;90 ms&lt;/td&gt;
&lt;td&gt;105 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Every number here includes my network path to the server and the keyring that attaches the key, which is why nothing is near the single-digit milliseconds the handler itself probably takes: a bare TLS handshake to the panel from the same worker took 65 and 97 ms in two tries. I can't see server time from outside, and the panel's own request log (&lt;code&gt;GET /api/requests/summary&lt;/code&gt;, added in v0.1.29) is owner-only, so I don't have p50 figures for the handlers. What the table does show is the comparison: the two endpoints that need container states are not slower than the one that only returns a version string, which is what you would expect if no Docker call is left on that path.&lt;/p&gt;

&lt;p&gt;What it doesn't show: the same calls before v0.1.28 (that release is long gone from the live panel), a long service list (my key sees one project, so &lt;code&gt;/api/services&lt;/code&gt; was 1,046 bytes), or the Docker usage endpoint, which needs a role on all projects that my key doesn't have.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this is the right trade for a deploy panel
&lt;/h2&gt;

&lt;p&gt;A deploy tool's own API is read far more than it is written. The panel polls, agents poll (Tugport's &lt;a href="https://tugport.com/docs/api/" rel="noopener noreferrer"&gt;MCP server&lt;/a&gt; and any script you write call the same endpoints), and a person with two tabs open is a poller too. Container state is already something the proxy needs to be right within seconds, because it routes traffic by it, so the API inherits an answer that was already being kept fresh for a harder job. Disk usage and release checks were never urgent.&lt;/p&gt;

&lt;p&gt;The price is staleness you can state: container state follows Docker events and is at most 30 seconds behind in the worst case, disk usage is at most about five minutes behind, and the update check is at most ten minutes behind. If you need an instant answer for one container, ask Docker; if you want the panel and your scripts to stay snappy while a build is eating the disk, this is the better default.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Does Tugport's API show live container state?&lt;/strong&gt;&lt;br&gt;
It shows the state from the proxy's latest container listing, which is rebuilt on container events (create, start, die, stop, pause, unpause, destroy, health changes) and every 30 seconds.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How fresh is the Docker disk usage in Tugport?&lt;/strong&gt;&lt;br&gt;
Up to five minutes old, refreshed in the background, and measured at startup so the first view is not slow.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which version has this?&lt;/strong&gt;&lt;br&gt;
v0.1.28 and everything after it. The live panel was on v0.1.35 when I tested.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://tugport.com/" rel="noopener noreferrer"&gt;Tugport&lt;/a&gt; is &lt;a href="https://github.com/backmeupplz/tugport" rel="noopener noreferrer"&gt;open source and MIT licensed&lt;/a&gt;; the &lt;a href="https://tugport.com/docs/" rel="noopener noreferrer"&gt;docs&lt;/a&gt; and the &lt;a href="https://tugport.com/docs/compare/" rel="noopener noreferrer"&gt;comparison with Coolify, Dokploy, Easypanel and CapRover&lt;/a&gt; are on the site. Facts in this post were read from the repository at commit &lt;code&gt;3166248&lt;/code&gt; and measured on 14 November 2026.&lt;/p&gt;

</description>
      <category>selfhosted</category>
      <category>docker</category>
      <category>rust</category>
      <category>devops</category>
    </item>
    <item>
      <title>Two weeks of Apple Search Ads on Todorant: 1,927 impressions, 16 taps and 3 installs for CA$13.97</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Fri, 13 Nov 2026 10:30:22 +0000</pubDate>
      <link>https://dev.to/backmeupplz/two-weeks-of-apple-search-ads-on-todorant-1927-impressions-16-taps-and-3-installs-for-ca1397-1ahp</link>
      <guid>https://dev.to/backmeupplz/two-weeks-of-apple-search-ads-on-todorant-1927-impressions-16-taps-and-3-installs-for-ca1397-1ahp</guid>
      <description>&lt;p&gt;Todorant's first Apple Search Ads test ran for 14 days, from October 29 to November 11, 2026, in the US App Store. It produced 1,927 impressions, 16 taps and 3 installs for CA$13.97, about US$10, which is far less than the CA$80 cap I had set. The eight exact-match keywords I had picked from store research got 105 of those impressions between them, and none reached 8 taps, which was my bar for calling a keyword a winner. The three installs all came from the Search Match campaign, whose daily budget was a quarter of the exact campaign's. This post has every number from Apple's report, what I changed halfway, and what I am trying in round two.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did the Apple Search Ads test for Todorant look like?
&lt;/h2&gt;

&lt;p&gt;I wanted to know which English keywords turn into installs before leaning on them in the App Store metadata. Todorant is a to-do manager that shows one task at a time, and its US listing is called "Todorant: To-Do List &amp;amp; Planner" with 13 ratings at 4.08 stars (App Store lookup, November 13, 2026).&lt;/p&gt;

&lt;p&gt;The account bills in Canadian dollars, so every amount below is CAD unless I say otherwise. There were two campaigns in the United States storefront:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Campaign&lt;/th&gt;
&lt;th&gt;Daily budget&lt;/th&gt;
&lt;th&gt;What it bought&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Exact&lt;/td&gt;
&lt;td&gt;CA$4&lt;/td&gt;
&lt;td&gt;eight keywords, exact match&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Discovery&lt;/td&gt;
&lt;td&gt;CA$1&lt;/td&gt;
&lt;td&gt;Search Match, no keywords, with the eight keywords and a few competitors' names as negatives&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The eight keywords were procrastination, eat the frog, to do list, todo list, daily planner, task manager, adhd planner and to do list widget, with max bids from CA$1.10 to CA$2.07. I had written down the rules before launch: a keyword needs at least 8 taps, a 40% tap-to-install rate and a cost per install of at most US$2.50 to count as a winner, and under 8 taps means "we can't tell". The exact campaign also rotated three custom product pages (beat procrastination, one task at a time, ADHD-friendly planner) against the default page, one per day.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did 14 days of Apple Search Ads actually deliver?
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Campaign&lt;/th&gt;
&lt;th&gt;Impressions&lt;/th&gt;
&lt;th&gt;Taps&lt;/th&gt;
&lt;th&gt;Installs&lt;/th&gt;
&lt;th&gt;Spend&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Exact&lt;/td&gt;
&lt;td&gt;105&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;CA$3.52&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Discovery&lt;/td&gt;
&lt;td&gt;1,822&lt;/td&gt;
&lt;td&gt;14&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;CA$10.45&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Total&lt;/td&gt;
&lt;td&gt;1,927&lt;/td&gt;
&lt;td&gt;16&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;CA$13.97&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Those are Apple's own numbers, pulled through the Campaign Management API for October 29 to November 11. The tap-through rate was 1.9% on the exact campaign and 0.8% on discovery. Of the 16 taps, 3 became installs, which is 19%, and the discovery campaign paid CA$3.48 per install (about US$2.51 at the 0.72 rate the ledger uses). That is right at my US$2.50 limit, but it comes from three installs, so I would not build anything on it.&lt;/p&gt;

&lt;p&gt;The budget was never the limit. The two campaigns could have spent CA$70 over the 14 days and spent about a fifth of that, because there were simply not enough auctions we entered.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why did the exact keywords get so few impressions?
&lt;/h2&gt;

&lt;p&gt;The 105 exact impressions split like this:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Keyword&lt;/th&gt;
&lt;th&gt;Max bid&lt;/th&gt;
&lt;th&gt;Impressions&lt;/th&gt;
&lt;th&gt;Taps&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;todo list&lt;/td&gt;
&lt;td&gt;CA$2.07&lt;/td&gt;
&lt;td&gt;46&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;to do list&lt;/td&gt;
&lt;td&gt;CA$2.07&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;to do list widget&lt;/td&gt;
&lt;td&gt;CA$1.68&lt;/td&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;adhd planner&lt;/td&gt;
&lt;td&gt;CA$1.95&lt;/td&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;task manager&lt;/td&gt;
&lt;td&gt;CA$1.79&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;eat the frog&lt;/td&gt;
&lt;td&gt;CA$1.10&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;procrastination&lt;/td&gt;
&lt;td&gt;CA$1.95&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;daily planner&lt;/td&gt;
&lt;td&gt;CA$2.09&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Two things stand out. The first is that "to do list" had 23 impressions on the first day and 7 on the second, then none for the remaining 12 days, even though the bid never changed. I can't see why Apple stopped showing the ad, and I do not want to guess at its auction.&lt;/p&gt;

&lt;p&gt;The second is the bid change. The plan said that after three full days, any keyword with no impressions would get CA$0.30 more, up to the CA$2.50 ceiling my API key allows. On November 1 that applied to procrastination, daily planner, adhd planner and to do list widget. The widget keyword got its first impression the next day, and adhd planner its first four days later, on November 5. Procrastination and daily planner stayed at zero with bids of CA$1.95 and CA$2.09. A keyword that is still unseen at the highest bid I am allowed to set is "too expensive", and I had written that down as a valid result.&lt;/p&gt;

&lt;p&gt;What I can show from outside Apple is the competition on the same searches. I ran the App Store search API for the US storefront today, November 13, 2026 (the same method as the October 27 research), and Todorant is not among the results for any of these searches:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Search&lt;/th&gt;
&lt;th&gt;Results&lt;/th&gt;
&lt;th&gt;Top result's ratings&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;to do list&lt;/td&gt;
&lt;td&gt;93&lt;/td&gt;
&lt;td&gt;49,154&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;todo list&lt;/td&gt;
&lt;td&gt;91&lt;/td&gt;
&lt;td&gt;49,154&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;adhd planner&lt;/td&gt;
&lt;td&gt;95&lt;/td&gt;
&lt;td&gt;168,839&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;to do list widget&lt;/td&gt;
&lt;td&gt;93&lt;/td&gt;
&lt;td&gt;49,154&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;procrastination&lt;/td&gt;
&lt;td&gt;92&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;eat the frog&lt;/td&gt;
&lt;td&gt;74&lt;/td&gt;
&lt;td&gt;28&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;With 13 ratings, Todorant is up against apps with tens of thousands for the generic words. For "procrastination" and "eat the frog" the leaders have between 0 and 28 ratings, so those two are where a small app could rank organically, which is a different question from what a tap costs.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did Search Match find, and what was noise?
&lt;/h2&gt;

&lt;p&gt;Search Match is where all three installs came from, and it is also where the budget went to words that have nothing to do with a to-do manager. Apple reported 21 search-term rows, and one of them, "(low volume)", is a bucket it does not break down. That bucket held 1,098 of the 1,822 impressions (60%), 8 taps and 2 installs. I cannot tell which searches those installs came from.&lt;/p&gt;

&lt;p&gt;The named terms were mostly work-scheduling and remote-desktop searches: workday (84 impressions), qgenda (69), todesk (68), proton calendar (49), workjam (39). Four terms got taps: "workbuddy" (2), "do" (2), a Chinese countdown-app term (1) and "involve" (1). Only "involve" came with an install, from 14 impressions and a CA$0.49 tap. That is a single event, and I treat it as noise rather than as a keyword to buy.&lt;/p&gt;

&lt;p&gt;Search Match volume also faded. The discovery campaign had 1,252 impressions in the first seven days (October 29 to November 4) and 570 in the second seven. I did not change that campaign in between, and I do not know why it shrank.&lt;/p&gt;

&lt;h2&gt;
  
  
  Did the custom product pages make a difference?
&lt;/h2&gt;

&lt;p&gt;Only the exact campaign rotated custom product pages, and it had 105 impressions in total. The default page had 51 impressions and no taps, the ADHD-friendly planner page 31 impressions and 1 tap, beat procrastination 11 impressions and 1 tap, and one task at a time 12 impressions and none. Two taps in total cannot separate one page from another, so I am reporting these only so nobody has to wonder whether I looked.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did I learn about keyword testing for a small app?
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Start with Search Match, not with your keyword list.&lt;/strong&gt; It was the only part that got volume, and the only thing it could show was how little of that volume looked like a to-do list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write the verdict rules before the test.&lt;/strong&gt; Because I had set "8 taps" in advance, the result was easy to state: no keyword qualified, and the metadata stays on the research-based words instead of picking a favorite from two taps.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Count how much of a report is hidden.&lt;/strong&gt; Sixty percent of the discovery impressions sat in Apple's "(low volume)" row, so any story about "which search term converted" would have covered a minority of the data.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A daily budget is a ceiling, not a purchase.&lt;/strong&gt; I budgeted CA$70 and spent CA$13.97. If you are planning a test around a number of taps, plan around the impressions you can win at your bids.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What happens in round two?
&lt;/h2&gt;

&lt;p&gt;Round two started today, November 13, and runs until November 20 with the same daily budgets (CA$4 and CA$1). It moves the four keywords that showed any impressions (to do list, todo list, adhd planner and to do list widget) from exact to broad match at the same bids, adds "involve" as an exact keyword, and gives the Search Match group "involve" as a negative. The winner rule is the one above: 8 taps, a 40% tap-to-install rate and a cost per install of at most US$2.50. I will write up the result after the week is over, whichever way it goes.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Data: Apple Search Ads Campaign Management API reports for Todorant's two US campaigns, October 29 to November 11, 2026, in CAD; App Store search and lookup APIs, US storefront, checked November 13, 2026. &lt;a href="https://todorant.com" rel="noopener noreferrer"&gt;Todorant&lt;/a&gt; is a to-do manager that shows one task at a time.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ios</category>
      <category>marketing</category>
      <category>appstore</category>
      <category>analytics</category>
    </item>
    <item>
      <title>I raced 20 AI agents for one kanban ticket, 50 times: how agentboard keeps exactly one winner</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Thu, 12 Nov 2026 19:00:25 +0000</pubDate>
      <link>https://dev.to/backmeupplz/i-raced-20-ai-agents-for-one-kanban-ticket-50-times-how-agentboard-keeps-exactly-one-winner-4hag</link>
      <guid>https://dev.to/backmeupplz/i-raced-20-ai-agents-for-one-kanban-ticket-50-times-how-agentboard-keeps-exactly-one-winner-4hag</guid>
      <description>&lt;p&gt;To keep several AI coding agents from picking the same ticket, make the claim a compare-and-set: the agent moves the ticket out of &lt;code&gt;to-do&lt;/code&gt; only if it is still in &lt;code&gt;to-do&lt;/code&gt;, and the board answers 409 to everyone who comes second. On 12 November 2026 I tested that on &lt;a href="https://agentboard.win" rel="noopener noreferrer"&gt;agentboard&lt;/a&gt; 0.1.13, my self-hosted kanban board for agents, by sending 20 agents after the same ticket at once, 50 times. Every round had exactly one winner and 19 agents got a 409. Without the check, all 20 of them won.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem: two agents, one ticket, two pull requests
&lt;/h2&gt;

&lt;p&gt;An agent that wants work lists the tickets in &lt;code&gt;to-do&lt;/code&gt;, picks the top one and moves it to &lt;code&gt;in-progress&lt;/code&gt;. With one agent that is fine. With several, two of them can read the same list within the same second, and both move the same ticket. Nothing fails, both agents say they have it, and an hour later I have two pull requests for one fix. I wrote about how that happened on my own board in &lt;a href="https://blog.borodutch.com/my-ai-agents-closed-198-tickets-in-25-days/" rel="noopener noreferrer"&gt;the post about the 198 tickets my agents closed in 25 days&lt;/a&gt;, and the fix there was a single field on the update call. This post is the test I should have run before trusting it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The claim, as one HTTP call
&lt;/h2&gt;

&lt;p&gt;An agent claims a ticket with a PATCH that names the column it expects the ticket to be in:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -X PATCH -H &lt;span class="s2"&gt;"Authorization: Bearer $KEY"&lt;/span&gt; -H &lt;span class="s1"&gt;'Content-Type: application/json'&lt;/span&gt; \
  &lt;span class="nv"&gt;$URL&lt;/span&gt;/api/tickets/42 \
  -d &lt;span class="s1"&gt;'{"if_column":"to-do","column":"in-progress","assignee":"me","comment":"Picking this up."}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The move, the assignment and the comment all happen in that one call. If the ticket is no longer in &lt;code&gt;to-do&lt;/code&gt;, nothing is written and the response is a 409 with a message the agent can read: &lt;code&gt;ticket #42 is in "in-progress", not "to-do"&lt;/code&gt;. The command line client does the same with &lt;code&gt;kb set 42 --if-column to-do --column in-progress --assignee me&lt;/code&gt;. I ran it twice on a fresh board: the first call exited 0, the second printed &lt;code&gt;409: ticket #71 is in "in-progress", not "to-do"&lt;/code&gt; and exited 1, so a shell script or an agent's tool call can branch on the exit code.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the check can't be beaten: it's one synchronous function
&lt;/h2&gt;

&lt;p&gt;I read the code at commit &lt;code&gt;b5deb33&lt;/code&gt; (tag v0.1.13). The handler for the update, &lt;code&gt;updateTicket&lt;/code&gt; in &lt;code&gt;server.js&lt;/code&gt;, loads the ticket, compares its column with &lt;code&gt;if_column&lt;/code&gt;, and then writes the new column, the assignee and the event rows in one transaction. The server uses &lt;code&gt;node:sqlite&lt;/code&gt;, whose calls are synchronous, so there is no &lt;code&gt;await&lt;/code&gt; between the read and the write. Node runs one request handler at a time, which means a second claim can only start after the first one has finished writing, and by then it sees &lt;code&gt;in-progress&lt;/code&gt; and gets its 409.&lt;br&gt;
The check lives in the application and not in the SQL statement, so it holds for the way agentboard is meant to run: one server process that owns one SQLite file. I did not test two server processes pointed at the same database, and I would not run it that way.&lt;/p&gt;
&lt;h2&gt;
  
  
  The test: 20 agents, 50 rounds
&lt;/h2&gt;

&lt;p&gt;I started a new board (Node 22.23.4, an empty database), created 20 agent users with &lt;code&gt;POST /api/users&lt;/code&gt;, each with its own key, and wrote a short script. Each round creates a ticket in &lt;code&gt;to-do&lt;/code&gt;, sends 20 claims at the same moment with &lt;code&gt;Promise.all&lt;/code&gt;, one per agent, and then reads the ticket back. Every request carried &lt;code&gt;"if_column": "to-do"&lt;/code&gt; and a comment saying which agent sent it.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Claim style&lt;/th&gt;
&lt;th&gt;Rounds&lt;/th&gt;
&lt;th&gt;Results per round&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;With &lt;code&gt;if_column&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;50&lt;/td&gt;
&lt;td&gt;1 × 200, 19 × 409&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Without &lt;code&gt;if_column&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;20 × 200&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;With the check, all 50 rounds ended the same way: one 200, nineteen 409s, and the assignee on the ticket was the agent that got the 200. The 50 rounds, about a thousand claims in all, took 1,260 milliseconds including creating and reading back the tickets. The same agent won every round, which says more about the order my script sent requests in than about the board, so don't read fairness into it. The board promises one winner, not a random one.&lt;br&gt;
Without the check, which is how a plain "move it to in-progress" call behaves, all 20 agents got a 200 in each of the 20 rounds. In the first round the ticket ended up with 20 comments saying "claimed by agent-N" and with the last writer, agent-20, as its assignee. Every one of those twenty agents believed it owned the ticket. That is the duplicate pull request, reproduced in about a second.&lt;/p&gt;
&lt;h2&gt;
  
  
  What an agent does with a 409
&lt;/h2&gt;

&lt;p&gt;A 409 on a claim is not an error to retry. It means somebody else is on that ticket, so the agent goes back to the list and takes another one:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="k"&gt;for&lt;/span&gt; id &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="nv"&gt;$(&lt;/span&gt;kb ls --column to-do --assignee none &lt;span class="o"&gt;|&lt;/span&gt; jq -r &lt;span class="s1"&gt;'.[].id'&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="k"&gt;do&lt;/span&gt;
  kb &lt;span class="nb"&gt;set&lt;/span&gt; &lt;span class="s2"&gt;"$id"&lt;/span&gt; --if-column to-do --column in-progress --assignee me \
    --comment &lt;span class="s2"&gt;"Picking this up"&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;break&lt;/span&gt;
&lt;span class="k"&gt;done&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I describe the whole loop, including the handoff comment and the pull request link, in &lt;a href="https://agentboard.win/claude-code-kanban-board" rel="noopener noreferrer"&gt;the guide to giving Claude Code a shared kanban board through the API&lt;/a&gt;, with the commands Claude Code ran.&lt;/p&gt;

&lt;h2&gt;
  
  
  Create and claim in one call
&lt;/h2&gt;

&lt;p&gt;The compare-and-set protects tickets that already sit in &lt;code&gt;to-do&lt;/code&gt;. There is a second race it can't help with: an agent that creates a ticket for work it is about to do itself. If it creates the ticket in &lt;code&gt;to-do&lt;/code&gt; and moves it a moment later, a dispatcher watching that column can grab it in between. That is the bug that gave me two pull requests for one fix on my real board. The creation call accepts the column and the assignee, so the ticket never sits unassigned. On the fresh board I ran &lt;code&gt;kb new "Fix other thing" --column in-progress --assignee me&lt;/code&gt; and the ticket came back already in &lt;code&gt;in-progress&lt;/code&gt;, assigned to the agent that created it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to copy if you build your own board
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Put the expected state in the claim itself, and answer 409 with the actual state in the message.&lt;/li&gt;
&lt;li&gt;Do the check and the write in one step with no &lt;code&gt;await&lt;/code&gt; in between, or move the check into the SQL &lt;code&gt;WHERE&lt;/code&gt; clause if your server runs as several processes.&lt;/li&gt;
&lt;li&gt;Test it with the requests really arriving at once. A claim that works when you call it twice in a row proves nothing about two agents calling it together.&lt;/li&gt;
&lt;li&gt;Let agents create tickets already assigned, so no ticket sits in the queue waiting to be grabbed.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Run it yourself
&lt;/h2&gt;

&lt;p&gt;agentboard is MIT-licensed and the source is on &lt;a href="https://github.com/backmeupplz/agentboard" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt;. One command starts it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;docker run -d -p 3000:3000 -v agentboard:/data ghcr.io/backmeupplz/agentboard
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open the setup link printed in the logs, add an agent under Settings, People, and give it the board's URL and its key. The race script is a short Node file using &lt;code&gt;fetch&lt;/code&gt;, and the API it calls is documented at &lt;code&gt;GET /api&lt;/code&gt; on your own board. Everything above was run on 12 November 2026 against agentboard 0.1.13 on Node 22.23.4; the numbers come from one machine and a local server, so treat the timings as a rough size and the one-winner result as the point.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>opensource</category>
      <category>selfhosted</category>
    </item>
    <item>
      <title>How ABS+ keeps your Audiobookshelf listening position in sync, even offline</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Wed, 11 Nov 2026 10:30:11 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-abs-keeps-your-audiobookshelf-listening-position-in-sync-even-offline-1gd1</link>
      <guid>https://dev.to/backmeupplz/how-abs-keeps-your-audiobookshelf-listening-position-in-sync-even-offline-1gd1</guid>
      <description>&lt;p&gt;&lt;a href="https://absplus.app" rel="noopener noreferrer"&gt;ABS+&lt;/a&gt; saves your Audiobookshelf listening position on the phone first and sends it to your server afterwards, so a chapter you listened to in a tunnel or on a plane is not lost and does not overwrite what you heard on another device in the meantime. Each play, pause and finish is written to storage before any network request is made, the server is asked what it already has before the app writes anything, and a failed send is retried with a delay that grows from one second to five minutes and survives the app being killed. This post walks through how that works in the code as it stands in the repository on November 9, 2026 (the sync files are unchanged since the 1.2.4 release of October 28), and why each rule exists.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why syncing a listening position is harder than it looks
&lt;/h2&gt;

&lt;p&gt;Audiobookshelf stores one progress record per book, and per podcast episode, on your server. The record holds the current time, the duration, a finished flag and a &lt;code&gt;lastUpdate&lt;/code&gt; timestamp. An app that wants to keep it current sends a &lt;code&gt;PATCH&lt;/code&gt; to &lt;code&gt;/api/me/progress/&amp;lt;item&amp;gt;&lt;/code&gt; now and then. The naive version of that is a single line in a timer, and it fails in the places where audiobooks are actually listened to: the train with no signal, the phone that dies mid-chapter, the second phone that was left paused at yesterday's position.&lt;/p&gt;

&lt;p&gt;Three things made me write a proper replay queue instead of the single line:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The server can be unreachable for hours while the audio keeps playing from a download.&lt;/li&gt;
&lt;li&gt;Audiobookshelf has no conditional write, so the app cannot say "update this only if nobody changed it". Whatever it does to avoid clobbering a newer position has to be done by reading first.&lt;/li&gt;
&lt;li&gt;The server answers a &lt;code&gt;PATCH&lt;/code&gt; without returning the progress row, so the app does not learn the timestamp the server gave to its own write unless it asks again.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 1: write to the phone before anything touches the network
&lt;/h2&gt;

&lt;p&gt;The Android player calls &lt;code&gt;record&lt;/code&gt; on every play and pause event, when a title ends, and on a 20-second tick while audio is playing. &lt;code&gt;record&lt;/code&gt; in &lt;code&gt;ProgressSync.kt&lt;/code&gt; works synchronously: it builds the new position, commits it to the app's preferences with a blocking &lt;code&gt;commit()&lt;/code&gt; (not the asynchronous &lt;code&gt;apply()&lt;/code&gt;), and only then wakes the background worker. If the write to storage fails, the Android code raises an error ("Could not save listening progress") instead of carrying on as if it had saved, and the iPhone app shows the same message as a toast. The iPhone version, &lt;code&gt;Progress.swift&lt;/code&gt;, has the same rule in its &lt;code&gt;push&lt;/code&gt; function: the snapshot is encoded and written atomically to disk, and the replay task starts only if that write succeeded.&lt;/p&gt;

&lt;p&gt;The journal keeps one entry per server, account, recipient and title, and that entry holds only the latest value. Listening to the same chapter for an hour offline therefore does not produce a queue of 180 updates to replay later; it produces one entry that says where you are now. The entry never contains a password or a token, only the position, the timestamps and the identity of the account it belongs to.&lt;/p&gt;

&lt;p&gt;Every event also gets a timestamp that can only move forward: the larger of the phone's clock and one millisecond past the previous timestamp recorded for that title. If a phone's clock runs behind the server's, the first read from the server is stored as well, so the next offline event still lands after it rather than before it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: ask the server first, and let a newer position win
&lt;/h2&gt;

&lt;p&gt;When the worker runs it does not send the saved position straight away. It reads the server's current record for that title first. If the server's &lt;code&gt;lastUpdate&lt;/code&gt; is newer than the one on the phone, someone listened somewhere else after this event happened, and the server's position wins: the phone adopts it and drops its own pending entry. This is last-write-wins by timestamp, chosen on purpose because it matches what a person expects. The device you used most recently decides where the book resumes.&lt;/p&gt;

&lt;p&gt;There is one exception that took the most tests to get right. When the app sends a position and the connection drops before the answer arrives, the server may have stored the write while the app believes it failed. On the next read the app sees a record newer than its own and would wrongly conclude that another device had written it. To avoid that, the app saves the payload it is about to send (&lt;code&gt;sent&lt;/code&gt;) before it sends it. When the server's record matches that payload, the app recognises its own write, keeps the server's timestamp as an acknowledgement, and does not treat it as a conflict. The comment in the code states the remaining gap plainly: with no conditional write in Audiobookshelf, a write by another device between the app's read and its send cannot be made atomic from the client.&lt;/p&gt;

&lt;p&gt;After a successful &lt;code&gt;PATCH&lt;/code&gt; the worker reads the record once more, because the response carries no progress row, and only then marks the entry as delivered. If a newer local event arrived while the request was in flight, that entry stays pending and goes out on the next pass.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: retry with a delay that survives a restart
&lt;/h2&gt;

&lt;p&gt;When a request fails, the entry stays pending and gets a retry time. On Android the delay is one second, then two, four, eight and so on, up to 256 seconds, and from the tenth failed attempt on, five minutes. The iPhone code uses the same curve, &lt;code&gt;min(300, 2^attempts)&lt;/code&gt; seconds, with the attempt counter capped at 10. The retry time is written into the journal, not kept in memory, so killing the app does not reset it, and starting the app wakes the worker (on the iPhone so do coming to the foreground and getting a connection back) without anything having to be playing. The test &lt;code&gt;automaticStartupDrainsPersistedQueueWithoutPlayback&lt;/code&gt; covers that case: it queues a position, starts the sync without playing anything, and expects the queue to drain on its own.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 4: a finished book stays finished
&lt;/h2&gt;

&lt;p&gt;A comment in the iPhone code records that Audiobookshelf treats an explicit &lt;code&gt;isFinished: false&lt;/code&gt; in a write as "mark this unread" and discards the position that comes with it. For that reason the app leaves the field out of a write unless the title is finished. A second trap sits on the playback side. After the last chapter ends, the player sends late callbacks, such as a pause or a restore at position zero, and Audiobookshelf clears the completed state when the position moves back from the end. The app keeps the completed position through those passive callbacks and starts a re-read only when playback really starts again. The test &lt;code&gt;passiveCompletionSurvivesButDeliberateOfflineRereadResumesAfterRestart&lt;/code&gt; runs exactly this: it finishes a book and a podcast episode offline, sends the late zero-position callbacks, checks the title still counts as 100 percent, then plays 24 seconds and checks that the server ends up with 24 seconds and no finished flag.&lt;/p&gt;

&lt;h2&gt;
  
  
  Linked accounts and the resume choice
&lt;/h2&gt;

&lt;p&gt;If you have more than one account on the same server and link them for a title, ABS+ records progress for each linked account, one entry each, acknowledged and retried independently. The test that plays two titles offline and replays them expects four pending entries, two titles times two recipients, and expects all four to arrive once the connection returns. When you open a title, the app resumes from the most recent position it knows for your own account, and it offers another linked account's position only if that one is newer and more than 30 seconds apart. Offline it does not even ask the server, and uses the phone's copy.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would tell someone building this
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Write the event to storage before the network. Everything else in the design follows from that order.&lt;/li&gt;
&lt;li&gt;Keep the latest value per title, not a log of events. A queue of positions only has to get you to the last one.&lt;/li&gt;
&lt;li&gt;Record what you are about to send, so that your own successful write is never mistaken for someone else's.&lt;/li&gt;
&lt;li&gt;Persist the retry schedule. A backoff that lives in memory restarts from zero every time the system kills the app, which on a phone is often.&lt;/li&gt;
&lt;li&gt;Test the awkward interleavings by name. &lt;code&gt;ProgressSyncTest.kt&lt;/code&gt; has 23 tests, among them an update that arrives during an in-flight first write, a token rotation while a newer event arrives, and a logout during a token refresh.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I read these files and their tests for this post but did not run the test suite today, so the behaviour described is what the code and tests say, not a fresh measurement on a device.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;ABS+ is a free, open-source player for your own Audiobookshelf server, with no ads and no analytics. You can get it on &lt;a href="https://play.google.com/store/apps/details?id=com.borodutch.absplus" rel="noopener noreferrer"&gt;Google Play&lt;/a&gt;, the &lt;a href="https://apps.apple.com/app/id6818353284" rel="noopener noreferrer"&gt;App Store&lt;/a&gt; and F-Droid, and the sync code is in the &lt;a href="https://github.com/backmeupplz/absplus" rel="noopener noreferrer"&gt;repository&lt;/a&gt; under &lt;code&gt;ProgressSync.kt&lt;/code&gt; for Android and &lt;code&gt;Progress.swift&lt;/code&gt; for iPhone. More about the app is on &lt;a href="https://absplus.app" rel="noopener noreferrer"&gt;absplus.app&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Does ABS+ lose my position if I am offline for a whole book?
&lt;/h3&gt;

&lt;p&gt;No. Every event is saved on the phone first and the newest position for each title is sent when the server is reachable again.&lt;/p&gt;

&lt;h3&gt;
  
  
  What happens if I listen on two phones?
&lt;/h3&gt;

&lt;p&gt;The position with the newer timestamp wins. The app reads the server's record before writing and adopts a newer one instead of overwriting it.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does progress sync need an ABS+ account?
&lt;/h3&gt;

&lt;p&gt;No. ABS+ talks only to your own Audiobookshelf server, and the saved queue holds positions and timestamps, never credentials.&lt;/p&gt;

</description>
      <category>android</category>
      <category>ios</category>
      <category>opensource</category>
      <category>selfhosted</category>
    </item>
    <item>
      <title>How MyGround backs up app databases: six dumps, not file copies</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Tue, 10 Nov 2026 10:30:11 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-myground-backs-up-app-databases-six-dumps-not-file-copies-3g4p</link>
      <guid>https://dev.to/backmeupplz/how-myground-backs-up-app-databases-six-dumps-not-file-copies-3g4p</guid>
      <description>&lt;p&gt;MyGround does not copy a running database's files when it backs up an app. For six of its 40 catalog apps (Immich, Nextcloud, Paperless-ngx, Firefly III, Kaneo and Penpot) it asks the database container for a dump with &lt;code&gt;pg_dumpall&lt;/code&gt; or &lt;code&gt;mariadb-dump&lt;/code&gt;, puts that dump into the restic snapshot instead of the raw data directory, and replays it on restore. For the apps that keep their data as plain files or in SQLite, it copies the files. This post shows how that works in the code, which apps it covers, and which ones it doesn't.&lt;/p&gt;

&lt;p&gt;Everything below was read from the MyGround repository at commit &lt;code&gt;3209dcf&lt;/code&gt; (version 0.1.112) on 2026-11-10. I read the code and counted the catalog files; I did not run a dump and restore on a live server for this post, so treat the behavior descriptions as what the source says, not as a test report.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a database needs a dump instead of a file copy
&lt;/h2&gt;

&lt;p&gt;A backup tool that copies &lt;code&gt;/var/lib/postgresql/data&lt;/code&gt; while Postgres is running can capture files from different moments: a table file written before a transaction and the write-ahead log from after it. A logical dump asks the database itself for a consistent view and writes it as SQL, which any later version of the same database can read. That is the reason MyGround's catalog has a &lt;code&gt;db_dump&lt;/code&gt; block for the apps that have a SQL database.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 40-app catalog, counted
&lt;/h2&gt;

&lt;p&gt;Each app in MyGround is a pair of files in the &lt;code&gt;apps/&lt;/code&gt; directory: a Docker Compose file and a TOML manifest with its name, health check, storage volumes and any extra settings. &lt;code&gt;ls apps/*.toml | wc -l&lt;/code&gt; gives 40 on the commit above. Seven of them ship a SQL or document database container of their own. Six of those seven carry a &lt;code&gt;db_dump&lt;/code&gt; entry:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;App&lt;/th&gt;
&lt;th&gt;Dump command&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Immich&lt;/td&gt;
&lt;td&gt;&lt;code&gt;pg_dumpall -U postgres&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Paperless-ngx&lt;/td&gt;
&lt;td&gt;&lt;code&gt;pg_dumpall -U paperless&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Kaneo&lt;/td&gt;
&lt;td&gt;&lt;code&gt;pg_dumpall -U kaneo&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Penpot&lt;/td&gt;
&lt;td&gt;&lt;code&gt;pg_dumpall -U penpot&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Nextcloud&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;mariadb-dump&lt;/code&gt; of the &lt;code&gt;nextcloud&lt;/code&gt; database&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Firefly III&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;mariadb-dump&lt;/code&gt; of the &lt;code&gt;firefly&lt;/code&gt; database&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The seventh is Anytype, whose manifest describes one &lt;code&gt;data&lt;/code&gt; volume holding MongoDB, Redis, MinIO and the any-sync nodes together. It has no dump entry, so its files are copied as they are. Five more apps (Groceries, Agentboard, Karakeep, Nomad and Headscale) keep a SQLite file in their data volume and are copied as files too.&lt;/p&gt;

&lt;p&gt;You can check this on your own copy of the repository with &lt;code&gt;grep -l db_dump apps/*.toml&lt;/code&gt;, which lists exactly the six manifests above.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happens during a backup
&lt;/h2&gt;

&lt;p&gt;A backup job loops over the app's storage volumes. For a volume without a dump entry, MyGround starts a restic container, mounts the volume's host path read-only at &lt;code&gt;/data&lt;/code&gt; and runs &lt;code&gt;restic backup&lt;/code&gt; with the tag &lt;code&gt;&amp;lt;app&amp;gt;/&amp;lt;volume&amp;gt;&lt;/code&gt;. For a volume that has one, the code in &lt;code&gt;src/backup.rs&lt;/code&gt; does this instead:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It takes a process-wide lock, so only one database dump runs at a time.&lt;/li&gt;
&lt;li&gt;It runs &lt;code&gt;docker exec &amp;lt;container&amp;gt; sh -c "&amp;lt;command&amp;gt; &amp;gt; /tmp/&amp;lt;dump file&amp;gt;"&lt;/code&gt;, so the dump is written inside the database container.&lt;/li&gt;
&lt;li&gt;It copies the file out with &lt;code&gt;docker cp&lt;/code&gt; into the app's &lt;code&gt;dumps&lt;/code&gt; directory on the host, and removes the temporary file inside the container.&lt;/li&gt;
&lt;li&gt;It backs up that &lt;code&gt;dumps&lt;/code&gt; directory with restic under the volume's tag.&lt;/li&gt;
&lt;li&gt;It deletes the dump file from the host once restic finishes.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Two consequences follow. The snapshot for the Immich &lt;code&gt;postgres&lt;/code&gt; volume contains &lt;code&gt;immich_db_dump.sql&lt;/code&gt; and not the Postgres data directory, so it is small and doesn't depend on the Postgres on-disk format. And if the dump command fails, the job fails: the loop stops with the error and the job is marked failed, instead of quietly falling back to a copy of the raw files. A broken dump shows up as a failed job in the dashboard on the day it happens, which is when you want to know.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happens during a restore
&lt;/h2&gt;

&lt;p&gt;Restoring a database snapshot is a different path from restoring files. The restore reports its progress in phases named &lt;code&gt;extracting&lt;/code&gt;, &lt;code&gt;wiping&lt;/code&gt;, &lt;code&gt;importing&lt;/code&gt; and &lt;code&gt;restoring&lt;/code&gt;. MyGround first extracts the snapshot with restic into a temporary directory and looks for the dump file by name. Only then does it run the app's &lt;code&gt;wipe_command&lt;/code&gt;, which for Immich is &lt;code&gt;DROP DATABASE IF EXISTS immich;&lt;/code&gt;, copy the dump into the container and run the app's &lt;code&gt;restore_command&lt;/code&gt; (&lt;code&gt;psql -U postgres &amp;lt; /tmp/immich_db_dump.sql&lt;/code&gt; for Immich). All six database apps have both a wipe and a restore command. The wipe happens after the extraction, so a snapshot that can't be read fails before anything in the running database is touched. The documentation page asks for a confirmation before this, because it replaces all current data.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't cover
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Apps outside the six.&lt;/strong&gt; A SQLite file copied while the app is writing to it can in principle be caught mid-write. The catalog has no per-app dump for those five, and Anytype's MongoDB files are copied the same way. I haven't tested a restore of any of them, so I can't say how often that matters in practice.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Restore of a different database version.&lt;/strong&gt; The dumps are SQL, which travels across versions better than raw files, but I haven't tried restoring a Postgres 14 dump into a newer image.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I'd tell someone setting this up
&lt;/h2&gt;

&lt;p&gt;Turn on a backup job for each of the six database apps and look at the first snapshot in the Backups page: it should contain the &lt;code&gt;.sql&lt;/code&gt; file, not a data directory. If you run Anytype or one of the SQLite apps, stop the app before an important snapshot, or use a destination you can restore from on a spare machine first. The &lt;a href="https://myground.online/docs.html#backup-overview" rel="noopener noreferrer"&gt;backup documentation&lt;/a&gt; describes jobs, destinations and restore, and the &lt;a href="https://blog.myground.online/s3-glacier-deep-archive-backup-cost/" rel="noopener noreferrer"&gt;S3 Glacier Deep Archive post&lt;/a&gt; covers what a cheaper destination costs.&lt;/p&gt;

&lt;p&gt;MyGround is &lt;a href="https://github.com/backmeupplz/myground" rel="noopener noreferrer"&gt;open source&lt;/a&gt; and installs with one command on a Linux or macOS machine. You can browse the catalog and the 40 manifests on &lt;a href="https://myground.online/docs.html" rel="noopener noreferrer"&gt;myground.online&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>selfhosting</category>
      <category>backup</category>
      <category>docker</category>
      <category>postgres</category>
    </item>
    <item>
      <title>Half of Quilcast's Google impressions came on one day: what profile pages did and didn't do</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Mon, 09 Nov 2026 16:00:28 +0000</pubDate>
      <link>https://dev.to/backmeupplz/half-of-quilcasts-google-impressions-came-on-one-day-what-profile-pages-did-and-didnt-do-415d</link>
      <guid>https://dev.to/backmeupplz/half-of-quilcasts-google-impressions-came-on-one-day-what-profile-pages-did-and-didnt-do-415d</guid>
      <description>&lt;p&gt;Half of the 5,591 Google impressions that Search Console lists for &lt;a href="https://quilcast.com" rel="noopener noreferrer"&gt;Quilcast&lt;/a&gt;, my Farcaster client in the browser, arrived on a single day, 27 October 2026, and the queries behind them are strings like "xnxnxnx" and "zzzxxxmmm", not anything a person looking for a Farcaster client would type. I found that out twelve days after I shipped profile pages that search engines can read, when I went to check whether they had worked, and the honest answer is that the numbers cannot say yet. This is what I looked at, what the pages now do, and how I changed what I count as success.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I shipped on 28 October
&lt;/h2&gt;

&lt;p&gt;Nearly all of quilcast.com's Google traffic was people searching for Farcaster usernames and landing on &lt;code&gt;/u/&amp;lt;username&amp;gt;&lt;/code&gt;. Those pages ranked (average position 10.3 on the day I started) but showed a bare bio as the snippet, "on Quilcast" in the title, a generic share image, and an empty &lt;code&gt;&amp;lt;div id="app"&amp;gt;&lt;/code&gt; to any crawler that does not run JavaScript. Quilcast 0.32.0 (&lt;a href="https://git.borodutch.com/backmeupplz/quilcast/pulls/75" rel="noopener noreferrer"&gt;pull request 75&lt;/a&gt;, merged on 28 October at 04:43 UTC) changed that:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The title is &lt;code&gt;&amp;lt;Name&amp;gt; (@username) on Farcaster · Quilcast&lt;/code&gt;, and the description is the bio plus the follower count plus "Read @username's latest casts on Quilcast".&lt;/li&gt;
&lt;li&gt;Every profile has its own 1200×630 share image at &lt;code&gt;/og/u/&amp;lt;fid&amp;gt;.png&lt;/code&gt;, and &lt;code&gt;ProfilePage&lt;/code&gt; plus &lt;code&gt;Person&lt;/code&gt; structured data with the follower and following counts.&lt;/li&gt;
&lt;li&gt;The profile header and the latest ten casts are plain HTML inside the page, which the app clears the moment it starts.&lt;/li&gt;
&lt;li&gt;Unknown usernames answer 404, and the followers and following lists are &lt;code&gt;noindex&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I checked the result live on 9 November: &lt;code&gt;quilcast.com/u/dwr&lt;/code&gt; answers 200 with the title "Dan Romero (&lt;a class="mentioned-user" href="https://dev.to/dwr"&gt;@dwr&lt;/a&gt;) on Farcaster · Quilcast", the canonical link, the two structured-data blocks and ten links to casts, and a made-up username answers 404. The hypothesis I wrote down was that click-through would rise from 0.8% to 2% within four weeks.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Search Console says twelve days later
&lt;/h2&gt;

&lt;p&gt;These are the daily rows my collector saved from the Search Console API for quilcast.com (the data lags two to three days, so the last row is 8 November):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Day&lt;/th&gt;
&lt;th&gt;Impressions&lt;/th&gt;
&lt;th&gt;Clicks&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;27 Oct&lt;/td&gt;
&lt;td&gt;2,328&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;28 Oct&lt;/td&gt;
&lt;td&gt;225&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;29 Oct&lt;/td&gt;
&lt;td&gt;199&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;30 Oct&lt;/td&gt;
&lt;td&gt;85&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;31 Oct&lt;/td&gt;
&lt;td&gt;24&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1 Nov&lt;/td&gt;
&lt;td&gt;31&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2 to 8 Nov&lt;/td&gt;
&lt;td&gt;241 in total&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The 28-day total on 8 November is 5,591 impressions and 43 clicks, with an average position of 11.5. So 42% of the impressions and 20 of the 43 clicks belong to the day before the change, and the twelve days after it add up to 805 impressions by the daily rows (967 by the growth of the 28-day total) and six clicks. If I had only looked at the headline total I would have said profile pages brought 5,000 impressions.&lt;/p&gt;

&lt;p&gt;The queries tell the same story. The top five in the window are "xnxnxnx" (57 impressions), "zzzxxxmmm" (30), "0xpolygon" (2), "angelica666" (2) and "nengsr" (1), and the top page for the first is the Farcaster profile with that username. Underneath them sit searches for another company's site, which are not about Farcaster at all. Nobody in this list searched for a Farcaster client, which means click-through rate on these pages is measuring how often a person who typed a username clicked the result for that username, and that tells me nothing about whether Quilcast is easier to find.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two measurement traps I fell into
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The recent days are not final.&lt;/strong&gt; Search Console revises the last few days upward. Comparing each day's row with the increase of the 28-day total the next morning, the total grew by more than the row said on most days (for example 262 against 199 for 29 October). Any "impressions collapsed" reading from the last row is partly this lag, and I would have written a different and wrong post if I had stated the drop as fact.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;One spike can be half the window.&lt;/strong&gt; With a few thousand impressions in total, one unexplained day dominates every average and every click-through rate. I cannot tell what happened on 27 October, and I am not going to guess in public. Search Console does not show per-day queries in the table I save, so the cause is not recoverable from my data.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I do now
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;The goal metric for Quilcast is signed-in accounts, not impressions or click-through on profile pages. Profile pages stay indexed, because they are the best page for anyone who does search a username, but I do not count them as growth until a count of sign-ins says so.&lt;/li&gt;
&lt;li&gt;Every Search Console number I publish gets the day it ends on, and any window with a single day above a third of the total is reported with and without that day.&lt;/li&gt;
&lt;li&gt;The four-week check on 25 November reads clicks on &lt;code&gt;/u/&lt;/code&gt; pages excluding the 27 October spike, and I will publish what it shows whether or not it favours me.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If you run something with a small search footprint, the cheap version of this is a spreadsheet with one column of daily impressions, and a habit of sorting it before you read the total. &lt;a href="https://quilcast.com" rel="noopener noreferrer"&gt;Quilcast&lt;/a&gt; itself is open source and runs in any browser, and its profile pages are an easy way to see what a Farcaster account looks like to a crawler: open any &lt;code&gt;/u/&amp;lt;username&amp;gt;&lt;/code&gt; page and view the source.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Data: Search Console API for quilcast.com via Radiant's daily collector, days 27 October to 8 November 2026; Quilcast pull request 75; live checks on 9 November 2026.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>seo</category>
      <category>webdev</category>
      <category>opensource</category>
      <category>farcaster</category>
    </item>
    <item>
      <title>I gave my AI agent a Tugport key that sees one project: Tugport's MCP server, tested</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Sun, 08 Nov 2026 16:00:19 +0000</pubDate>
      <link>https://dev.to/backmeupplz/i-gave-my-ai-agent-a-tugport-key-that-sees-one-project-tugports-mcp-server-tested-2p24</link>
      <guid>https://dev.to/backmeupplz/i-gave-my-ai-agent-a-tugport-key-that-sees-one-project-tugports-mcp-server-tested-2p24</guid>
      <description>&lt;p&gt;Tugport 0.1.31 can run as an MCP server over stdio: an agent starts &lt;code&gt;tugport mcp&lt;/code&gt; as a command, and it gets 32 tools for projects, deploys, logs, env vars, domains, backups and the audit log, forwarded to your panel with the agent's own API key. I tested it on 8 November 2026 against the panel that runs most of my sites, using the key my own AI agent works with every day, and what stood out is how little that key can see: one project, with the operator role, and nothing else.&lt;/p&gt;

&lt;h2&gt;
  
  
  What &lt;code&gt;tugport mcp&lt;/code&gt; does
&lt;/h2&gt;

&lt;p&gt;MCP clients such as Claude Code and Codex either connect to a URL or start a command and talk to it over standard input and output. Tugport already had the URL kind, a &lt;code&gt;POST /mcp&lt;/code&gt; endpoint on the panel. Version 0.1.31, released on 29 October, added the command kind (&lt;a href="https://git.borodutch.com/backmeupplz/tugport/commit/6af9168a156ba913eeec3c6ac2c2643650b32824" rel="noopener noreferrer"&gt;commit 6af9168&lt;/a&gt;). The diff is about thirty lines in &lt;code&gt;src/cli.rs&lt;/code&gt;: it reads one JSON-RPC message per line, posts each to the panel's &lt;code&gt;/mcp&lt;/code&gt; with the saved bearer key, and writes the reply back as one line. A message that is only a notification gets a 202 and no output, which is what the MCP stdio transport expects.&lt;br&gt;
You point it at a panel the same way as the rest of the CLI, with &lt;code&gt;tugport login https://panel.example.com --token tpk_...&lt;/code&gt;, or with the &lt;code&gt;TUGPORT_URL&lt;/code&gt; and &lt;code&gt;TUGPORT_TOKEN&lt;/code&gt; environment variables inside the client's config:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"tugport"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"tugport"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"TUGPORT_URL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://panel.example.com"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"TUGPORT_TOKEN"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"tpk_..."&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The release binary is a single file and needs nothing else on the machine running the agent, so the agent does not have to live on the server it deploys to.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I ran, and what came back
&lt;/h2&gt;

&lt;p&gt;First the part that needs no panel at all. I downloaded &lt;code&gt;tugport-linux-x86_64&lt;/code&gt; from the v0.1.31 release, started &lt;code&gt;tugport mcp&lt;/code&gt; with no login, and sent it the three messages a client sends at startup: &lt;code&gt;initialize&lt;/code&gt;, the &lt;code&gt;initialized&lt;/code&gt; notification and &lt;code&gt;tools/list&lt;/code&gt;. It answered with protocol version 2025-06-18, server name &lt;code&gt;tugport&lt;/code&gt;, version 0.1.31 and a list of 32 tools. Asked to actually call one (&lt;code&gt;list_projects&lt;/code&gt;), it replied with a plain message that it is not connected to a panel and how to log in, flagged as an error. That offline mode is deliberate: the commit says it exists so MCP directories can inspect the server without credentials.&lt;br&gt;
Then the part with a panel. I sent the same &lt;code&gt;tools/list&lt;/code&gt; and a few tool calls straight to &lt;code&gt;https://admin.borodutch.com/mcp&lt;/code&gt; with the agent's key, since that is the endpoint the stdio command forwards to. The panel reported version 0.1.31 (commit 6af9168) and Docker 29.9.2. I have not run the stdio command against the panel itself; I read what it forwards in the source.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;list_projects&lt;/code&gt; returned exactly one project, &lt;code&gt;marketing&lt;/code&gt;, with my agent's role as &lt;code&gt;operator&lt;/code&gt;, seven services of which six were running and one stopped.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;get_service&lt;/code&gt; on &lt;code&gt;marketing.kuma-test&lt;/code&gt;, the one service my agent uses for tested guides, returned its domain, port 3001, the image &lt;code&gt;louislam/uptime-kuma:2&lt;/code&gt; and the last deploy with its actor (&lt;code&gt;radiant&lt;/code&gt;) and trigger (&lt;code&gt;agent&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;list_env&lt;/code&gt; returned &lt;code&gt;"values_visible": false&lt;/code&gt;: this key sees variable names, never values, because it has not been granted &lt;code&gt;secrets:read&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;get_project&lt;/code&gt; for &lt;code&gt;default&lt;/code&gt; and for &lt;code&gt;borodutch&lt;/code&gt; both answered &lt;code&gt;project not found&lt;/code&gt;. Those projects may or may not exist on my panel; the point is that to this key they look the same as projects that do not exist, so an agent cannot even learn their names.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key's identity from &lt;code&gt;/api/me&lt;/code&gt; says the same thing in one line: kind &lt;code&gt;agent&lt;/code&gt;, no all-projects role, no grants, a single project role.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 32 tools, grouped
&lt;/h2&gt;

&lt;p&gt;The list is wider than I expected from a deploy tool. By what they touch:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Projects and services:&lt;/strong&gt; &lt;code&gt;list_projects&lt;/code&gt;, &lt;code&gt;get_project&lt;/code&gt;, &lt;code&gt;create_project&lt;/code&gt;, &lt;code&gt;create_service&lt;/code&gt;, &lt;code&gt;get_service&lt;/code&gt;, &lt;code&gt;update_service&lt;/code&gt;, &lt;code&gt;delete_service&lt;/code&gt;, &lt;code&gt;list_services&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deploys:&lt;/strong&gt; &lt;code&gt;deploy&lt;/code&gt; (waits for the result unless told not to), &lt;code&gt;list_deploys&lt;/code&gt;, &lt;code&gt;get_deploy&lt;/code&gt; with the full log, &lt;code&gt;cancel_deploy&lt;/code&gt;, &lt;code&gt;rollback&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Running services:&lt;/strong&gt; &lt;code&gt;restart&lt;/code&gt;, &lt;code&gt;stop&lt;/code&gt;, &lt;code&gt;start&lt;/code&gt;, &lt;code&gt;get_logs&lt;/code&gt;, &lt;code&gt;exec&lt;/code&gt;, &lt;code&gt;list_env&lt;/code&gt;, &lt;code&gt;set_env&lt;/code&gt;, &lt;code&gt;add_domain&lt;/code&gt;, &lt;code&gt;remove_domain&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Servers and metrics:&lt;/strong&gt; &lt;code&gt;list_servers&lt;/code&gt;, &lt;code&gt;server_status&lt;/code&gt;, &lt;code&gt;metrics&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Databases and backups:&lt;/strong&gt; &lt;code&gt;list_replicas&lt;/code&gt;, &lt;code&gt;add_replica&lt;/code&gt;, &lt;code&gt;promote_replica&lt;/code&gt;, &lt;code&gt;list_backups&lt;/code&gt;, &lt;code&gt;run_backup&lt;/code&gt;, &lt;code&gt;backup_run&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audit:&lt;/strong&gt; &lt;code&gt;audit_log&lt;/code&gt;, who did what, newest first.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Putting &lt;code&gt;delete_service&lt;/code&gt; and &lt;code&gt;exec&lt;/code&gt; in front of an agent sounds alarming until you remember that the tool list is the same for everyone and the permissions are not. The roles from the API reference apply to every tool call: a viewer reads status, logs and variable names, a deployer deploys, restarts and rolls back, an operator changes settings and creates services, and an admin deletes things. Deleting also needs the separate &lt;code&gt;destroy&lt;/code&gt; grant, and seeing variable values needs &lt;code&gt;secrets:read&lt;/code&gt;. My agent has the operator role and neither grant, which is why it can create, deploy and stop its test service but cannot delete one, and why I do not have to trust it not to try.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why I prefer a key that sees one project
&lt;/h2&gt;

&lt;p&gt;The usual way to give an agent access to a server is a shell, which is everything or nothing. With a scoped Tugport key the agent sees a single project, can only do what its role allows, and every call it makes lands in the audit log under its own name, with the key that made it. If it misbehaves I revoke that one key with &lt;code&gt;DELETE /api/keys/{id}&lt;/code&gt; and nothing else changes. If I want to hand a second agent the same tools for a different project, I run &lt;code&gt;tugport agent &amp;lt;name&amp;gt; --project &amp;lt;project&amp;gt;:&amp;lt;role&amp;gt;&lt;/code&gt; on the panel, which prints the key once, and optionally add &lt;code&gt;--expires-days&lt;/code&gt; so it stops working by itself.&lt;br&gt;
The same setup works from any client that speaks MCP, so the instructions for one agent are the instructions for all of them. An agent that starts commands gets &lt;code&gt;tugport mcp&lt;/code&gt;; one that connects to URLs gets &lt;code&gt;/mcp&lt;/code&gt;; a script gets the REST API; and all three are held to the same roles.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I did not test
&lt;/h2&gt;

&lt;p&gt;I did not call &lt;code&gt;delete_service&lt;/code&gt;, &lt;code&gt;exec&lt;/code&gt;, &lt;code&gt;add_replica&lt;/code&gt; or any backup tool, because they change production and my agent's job on this panel does not include them. I also did not measure latency: every call I made went over the internet and through a proxy that holds the key, so any number I put here would describe that path, not Tugport.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;Install &lt;a href="https://tugport.com" rel="noopener noreferrer"&gt;Tugport&lt;/a&gt; on a VPS with &lt;code&gt;curl -fsSL https://tugport.com/install | sh&lt;/code&gt;, open the one-time link it prints to create the owner account, create an agent user for one project, and give its key to your client. The &lt;a href="https://tugport.com/docs/api/" rel="noopener noreferrer"&gt;API reference&lt;/a&gt; lists every endpoint and the MCP section, and the &lt;a href="https://github.com/backmeupplz/tugport" rel="noopener noreferrer"&gt;repository&lt;/a&gt; has the source for everything above. Tugport is young (0.1.x), so expect rough edges and tell me about them.&lt;br&gt;
&lt;em&gt;Tested on 8 November 2026 with Tugport 0.1.31 (linux x86_64 client and the panel at 0.1.31, commit 6af9168).&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>ai</category>
      <category>selfhosted</category>
      <category>devops</category>
    </item>
    <item>
      <title>A button beat a line of text six to one: what reply promos in my Telegram bots measured in ten days</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Sat, 07 Nov 2026 16:00:22 +0000</pubDate>
      <link>https://dev.to/backmeupplz/a-button-beat-a-line-of-text-six-to-one-what-reply-promos-in-my-telegram-bots-measured-in-ten-days-5836</link>
      <guid>https://dev.to/backmeupplz/a-button-beat-a-line-of-text-six-to-one-what-reply-promos-in-my-telegram-bots-measured-in-ten-days-5836</guid>
      <description>&lt;p&gt;A button under a bot's reply brought people to my other bots at six times the rate of a line of text saying the same thing. Across 784 people shown a promo in my Telegram bots between 28 October and 7 November 2026, 43 of 397 who got a button opened the other bot (10.8%), against 7 of 387 who got a caption line (1.8%). I did not expect the gap to be that wide, and the way the safety rule behaved during the tests taught me more than the winner did.&lt;/p&gt;

&lt;p&gt;The bots are &lt;a href="https://voicybot.com" rel="noopener noreferrer"&gt;Voicy, which turns voice messages into text&lt;/a&gt;, &lt;a href="https://jev.borodutch.com" rel="noopener noreferrer"&gt;Jev Antispam&lt;/a&gt;, &lt;a href="https://randy.borodutch.com" rel="noopener noreferrer"&gt;Randy&lt;/a&gt;, &lt;a href="https://banofbot.borodutch.com" rel="noopener noreferrer"&gt;Banofbot&lt;/a&gt;, Temply and &lt;a href="https://videodl.borodutch.com" rel="noopener noreferrer"&gt;VideoDownloadBot&lt;/a&gt;. All numbers below come from the reports my own tooling printed on 7 November 2026 (&lt;code&gt;python3 bin/promos.py report&lt;/code&gt; and the log of each experiment).&lt;/p&gt;

&lt;h2&gt;
  
  
  How a reply promo test is set up
&lt;/h2&gt;

&lt;p&gt;Every bot carries the same small hook. When a bot is about to answer a person in a private chat, the hook hashes the campaign name together with the Telegram user id and uses the result to put that person in one of three arms: a caption line added to the reply, a URL button under the reply, or a holdout that never sees anything. The same person always lands in the same arm, so nobody sees a promo, loses it and sees it again. Seven bots have the hook live today.&lt;/p&gt;

&lt;p&gt;The rules were the same in every test. Private chats only, one promo per reply, at most one per person every 14 days, and every text says it comes from the same maker. The line reads something like "From the makers of this bot: Jev Antispam quietly deletes spam in your groups". The button says "Jev: antispam for groups". Both point to a &lt;code&gt;t.me/&amp;lt;bot&amp;gt;?start=&amp;lt;campaign&amp;gt;-&amp;lt;variant&amp;gt;-&amp;lt;host code&amp;gt;&lt;/code&gt; deep link, so Telegram opens the other bot directly and that bot can tell which campaign and arm sent the person.&lt;/p&gt;

&lt;p&gt;Three numbers matter. A &lt;strong&gt;shown&lt;/strong&gt; person had the promo rendered in a reply. An &lt;strong&gt;arrival&lt;/strong&gt; is a shown person who then pressed Start on the target bot with that payload. A &lt;strong&gt;conversion&lt;/strong&gt; is an arrival who then actually used the target bot within seven days. The holdout is there for one reason: it tells me how many people would have shown up anyway, and how many would have blocked the bot anyway.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the numbers say: the button wins, by a lot
&lt;/h2&gt;

&lt;p&gt;Seven tests had shown both a line and a button to anyone, and I also ran one by hand (VideoDownloadBot pointing to Voicy). Pooled:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Arm&lt;/th&gt;
&lt;th&gt;Shown&lt;/th&gt;
&lt;th&gt;Arrivals&lt;/th&gt;
&lt;th&gt;Rate&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Caption line&lt;/td&gt;
&lt;td&gt;387&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;1.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;URL button&lt;/td&gt;
&lt;td&gt;397&lt;/td&gt;
&lt;td&gt;43&lt;/td&gt;
&lt;td&gt;10.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Holdout&lt;/td&gt;
&lt;td&gt;no promo&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Fifty arrivals in total, and six conversions: four of them came from one button test, two from caption lines. The gap is not a fluke of pooling. Inside every test the arms were assigned by hash, so each test compares like with like, and a one-sided exact test on the pooled counts gives a p-value below 0.0001.&lt;/p&gt;

&lt;p&gt;The three tests hosted by Voicy, which has by far the most traffic, show it best:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Test&lt;/th&gt;
&lt;th&gt;Line&lt;/th&gt;
&lt;th&gt;Button&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Voicy to VideoDownloadBot, "free" wording&lt;/td&gt;
&lt;td&gt;2 of 60&lt;/td&gt;
&lt;td&gt;13 of 42&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Voicy to VideoDownloadBot, "what it does" wording&lt;/td&gt;
&lt;td&gt;2 of 54&lt;/td&gt;
&lt;td&gt;4 of 50&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Voicy to Jev Antispam&lt;/td&gt;
&lt;td&gt;1 of 36&lt;/td&gt;
&lt;td&gt;15 of 68&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Together that is 5 arrivals from 150 people shown a line and 32 from 160 shown a button. The reverse direction agrees: VideoDownloadBot pointing to Voicy got 1 arrival from 175 line impressions and 5 from 173 button impressions. In the five smaller bots (Banofbot, Jev, Randy, Temply, Todorant) the samples are tiny, 1 of 61 for the line and 6 of 64 for the button, and one test, Banofbot to Jev, produced no arrivals at all from 45 people shown either version. I would not read anything into those yet.&lt;/p&gt;

&lt;p&gt;My best guess for why a button wins is boring: it is a thing you can press, it sits where your thumb already is, and a caption line is one more sentence in a reply you came for. I have not tested that guess, so treat it as a guess. What I did test is that the quiet version, the line, is almost indistinguishable from nothing.&lt;/p&gt;

&lt;h2&gt;
  
  
  The safety rule fired three times, and it was right once
&lt;/h2&gt;

&lt;p&gt;Each test has stop rules written down before it starts. The one that matters here watches new blocks: it counts people who blocked the bot after being counted in a test, in treatment and in the holdout, and stops the test and emails me if treatment is meaningfully worse. The first version of the rule had two clauses. The first stopped a test when treatment's block rate exceeded the holdout's by one point and a one-sided test said p below 0.1, after 100 people. The second stopped it when treatment's block rate reached 3%, no questions asked.&lt;/p&gt;

&lt;p&gt;That second clause was a mistake for Voicy, and I only saw it on the first test. About a quarter of the Voicy users in a test block the bot during the test window whether or not they ever saw a promo: in the Voicy-to-Jev test the holdout had 264 blocks among 1,121 people (23.6%). I have not worked out why. A 3% ceiling on a number that sits at 24 percent stops everything, every time.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;31 October, Voicy to VideoDownloadBot:&lt;/strong&gt; the test was stopped with 21.6% blocks in treatment (22 of 102). The holdout was at 23% (208 of 905). Treatment was actually lower, so this was a false alarm. I changed the clause to also require treatment to be a point above the holdout.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;3 November, Voicy to VideoDownloadBot again, with new wording:&lt;/strong&gt; stopped with 35.6% blocks in treatment (37 of 104) against 26.2% in the holdout. A gap of nine points on a sample that size is about two standard errors, and I treated it as real. The test ended there. It brought 6 arrivals from 104 people, so the traffic did not justify finding out.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;7 November, Voicy to Jev Antispam:&lt;/strong&gt; stopped at 25.2% (26 of 103) against 23.6%. That is 1.6 points over the holdout, enough to trip the "three percent and one point over" clause, and well inside noise for 103 people. The button had just got 15 arrivals from 68 people. Another false alarm.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the rule did what I wanted exactly once, the second time, and the other two stops cost me a few days of data each. The fix I take from this is not a looser number. It is that the outright clause needs the same significance check as the first one, because a safety rule with an absolute threshold only works when the baseline is far below it. The code does not check significance on that clause yet, so until it does, a Voicy-hosted test can stop early on noise.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I changed because of this
&lt;/h2&gt;

&lt;p&gt;Caption lines are retired. New tests are button against button, with two wordings or two targets, because a test that includes the line spends 5% of a bot's users on an arm I already know is weak. The first of those, Voicy to Randy, started on 7 November, with a button and a 95% holdout, and has shown a promo to one person so far.&lt;/p&gt;

&lt;p&gt;The holdout stays large: 80 to 90% of a bot's users in these tests, and never below 50% as a rule. Without it I would not know that a quarter of Voicy's test users block the bot on their own, and I could not have told the 31 October stop (treatment no worse than the holdout) from the 3 November one (nine points worse).&lt;/p&gt;

&lt;h2&gt;
  
  
  What you can copy
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Use a button, not a sentence. If your bot's reply already carries a keyboard, adding a button costs one line of code.&lt;/li&gt;
&lt;li&gt;Keep a holdout that never sees the promo, and compare blocks against the holdout, not against zero.&lt;/li&gt;
&lt;li&gt;Measure the baseline before you write a stop rule. If your normal block rate is 24%, a rule that says "stop at 3%" is not a safety rule, it is an off switch.&lt;/li&gt;
&lt;li&gt;Count arrivals and the use that follows separately. Fifty people pressed Start and six used the other bot within a week; that second number is the one that pays for the effort.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The tests are still running, one after another, and I will write down the next result when a button wins or loses against another button. If you want to see what these bots do before the promo shows up, &lt;a href="https://voicybot.com" rel="noopener noreferrer"&gt;Voicy&lt;/a&gt; is the one with the most traffic, and &lt;a href="https://t.me/videodownloadbot" rel="noopener noreferrer"&gt;VideoDownloadBot&lt;/a&gt; is the one that most Voicy users were pointed to.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Numbers read on 7 November 2026 from the bots' own event reports; arrivals are people who pressed Start on the target bot, conversions are people who used it within seven days. Tests started between 28 October and 7 November, so older ones have had more time to collect results.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>telegram</category>
      <category>bots</category>
      <category>testing</category>
      <category>indiehackers</category>
    </item>
    <item>
      <title>How AI agents sign in to a kanban board: users with API keys, not an integration</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Fri, 06 Nov 2026 10:30:23 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-ai-agents-sign-in-to-a-kanban-board-users-with-api-keys-not-an-integration-4c4c</link>
      <guid>https://dev.to/backmeupplz/how-ai-agents-sign-in-to-a-kanban-board-users-with-api-keys-not-an-integration-4c4c</guid>
      <description>&lt;p&gt;In &lt;a href="https://agentboard.win" rel="noopener noreferrer"&gt;agentboard&lt;/a&gt;, an AI agent is a user with an API key, and that is the whole integration. A human creates the agent under Settings → People, the server shows a key that starts with &lt;code&gt;ab_&lt;/code&gt; exactly once, and the agent sends it as &lt;code&gt;Authorization: Bearer ab_…&lt;/code&gt; on every call. There is no OAuth app, no plugin and no MCP server in between. I read the auth code in the repository today (version 0.1.13, commit 22ca453), started the server on my machine and ran every claim below against it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is in an agent key
&lt;/h2&gt;

&lt;p&gt;A key is &lt;code&gt;ab_&lt;/code&gt; plus 24 random bytes in base64url, which is 35 characters in total. The server never stores it. It stores the SHA-256 of the key in a &lt;code&gt;key_hash&lt;/code&gt; column, and on each request it hashes the bearer token and looks that hash up. A fast hash is fine here because the input is 192 random bits rather than a password somebody chose, so there is nothing to brute-force.&lt;br&gt;
Humans are different. Their passwords go through scrypt with a 16-byte random salt, and the login route runs scrypt even when the email doesn't exist (against a dummy hash), so the response time doesn't tell an attacker which emails have accounts. Failed sign-ins are limited to 10 per IP in 15 minutes; I sent 10 wrong passwords and the eleventh request came back &lt;code&gt;429&lt;/code&gt;. The limiter lives in memory and resets when the server restarts, which is acceptable for a board that runs as one process.&lt;br&gt;
The browser gets a 32-byte session token in an &lt;code&gt;HttpOnly&lt;/code&gt;, &lt;code&gt;SameSite=Lax&lt;/code&gt; cookie that lasts 90 days, and again only its SHA-256 is stored. Writes that carry an &lt;code&gt;Origin&lt;/code&gt; header from another site are refused, because on a shared domain such as &lt;code&gt;*.ts.net&lt;/code&gt; other hosts count as same-site and the cookie setting alone doesn't protect you. I tested that too: a cookie-authenticated write with &lt;code&gt;Origin: https://evil.example&lt;/code&gt; returned &lt;code&gt;403 cross-origin request refused&lt;/code&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  What an agent can and cannot do
&lt;/h2&gt;

&lt;p&gt;Every user row has a &lt;code&gt;kind&lt;/code&gt;, either &lt;code&gt;human&lt;/code&gt; or &lt;code&gt;agent&lt;/code&gt;, and the server enforces it per route. Agents get everything that is work: create, read, move, assign and comment on tickets, upload files, manage projects, read the event feed and open the live stream. Anything that changes who can use the board or how it is laid out is humans only. Using the key I had just created, I tried four of those calls:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;POST /api/users&lt;/code&gt; (create another user) returned &lt;code&gt;403 humans only&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;POST /api/users/astra/key&lt;/code&gt; (issue a key, even for itself) returned &lt;code&gt;403 humans only&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;POST /api/columns&lt;/code&gt; returned &lt;code&gt;403 humans only&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;DELETE /api/tickets/1&lt;/code&gt; returned &lt;code&gt;403 humans only&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I like this split because the failure mode is boring. An agent that goes off the rails can fill the board with bad tickets, which a human can clean up, but it cannot hand itself more access, add accomplices or delete the history that shows what it did. Agents also don't see other people's email addresses: &lt;code&gt;GET /api/users&lt;/code&gt; returns names, kinds and whether a key exists, and the email field only appears for human callers.&lt;/p&gt;
&lt;h2&gt;
  
  
  How two agents avoid grabbing the same ticket
&lt;/h2&gt;

&lt;p&gt;Authentication tells the server who is calling. The next problem with several agents is two of them picking up the same ticket. Agentboard handles it with a compare-and-set field on the update call, &lt;code&gt;if_column&lt;/code&gt;. The claim looks like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-X&lt;/span&gt; PATCH &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Authorization: Bearer &lt;/span&gt;&lt;span class="nv"&gt;$KEY&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s1"&gt;'Content-Type: application/json'&lt;/span&gt; &lt;span class="nv"&gt;$URL&lt;/span&gt;/api/tickets/42 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{"if_column":"to-do","column":"in-progress","assignee":"me","comment":"Picking this up."}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The move, the assignment and the comment happen in one call. I ran it twice against the same ticket: the first call returned &lt;code&gt;200&lt;/code&gt;, and the second one returned &lt;code&gt;409&lt;/code&gt; with &lt;code&gt;ticket #1 is in "in-progress", not "to-do"&lt;/code&gt;. The loser of the race learns it lost from the status code and goes to the next ticket, with no locks and no coordinator process. &lt;code&gt;"assignee": "me"&lt;/code&gt; resolves to whichever user owns the key, so the same prompt works for every agent.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rotating a key
&lt;/h2&gt;

&lt;p&gt;A person with a session issues a new key with &lt;code&gt;POST /api/users/&amp;lt;name&amp;gt;/key&lt;/code&gt;. The old key stops working at once: after I rotated astra's key, the old one got &lt;code&gt;401&lt;/code&gt; and the new one got &lt;code&gt;200&lt;/code&gt; on &lt;code&gt;/api/me&lt;/code&gt;. The server also closes every live event stream that was opened with the old key (the code does the same when a user is deleted or a session signs out, which I didn't run). I opened a stream with the old key and rotated; the connection ended after about one second instead of staying open and silently listening on a revoked credential. That detail matters more than it sounds, because an agent that watches the board for hours would otherwise keep receiving events after you thought you had cut it off.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not OAuth or an MCP server
&lt;/h2&gt;

&lt;p&gt;The one thing every agent runtime does reliably is send an HTTP header. A bearer key needs nothing from the agent's runtime: Claude Code, Codex, a cron job and a shell script all use the same &lt;code&gt;curl&lt;/code&gt; line, and the server serves its own documentation at &lt;code&gt;GET /api&lt;/code&gt; as markdown so an agent can read how to use it from the board itself. The bundled zero-dependency CLI, &lt;code&gt;kb&lt;/code&gt;, reads &lt;code&gt;AGENTBOARD_URL&lt;/code&gt; and &lt;code&gt;AGENTBOARD_KEY&lt;/code&gt; and nothing else. The cost is that a key is a long-lived secret, so treat it like one: give each agent its own key, so a leak or a misbehaving agent is one rotation and not a redeploy for everyone.&lt;br&gt;
There is also an attribution benefit. Every event on a ticket records the user and whether it was an agent or a human, so the activity feed answers who moved what without any extra logging. The &lt;a href="https://blog.borodutch.com/my-ai-agents-closed-198-tickets-in-25-days/" rel="noopener noreferrer"&gt;post about 198 tickets closed in 25 days&lt;/a&gt; shows what that history looks like on a real board.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it yourself
&lt;/h2&gt;

&lt;p&gt;The server has no dependencies: Node's built-in &lt;code&gt;node:http&lt;/code&gt; and &lt;code&gt;node:sqlite&lt;/code&gt;. It declares Node 24 or newer, and I ran today's tests on Node 22.23.4 without trouble. Set &lt;code&gt;ADMIN_EMAIL&lt;/code&gt; and &lt;code&gt;ADMIN_PASSWORD&lt;/code&gt; for the first owner (or open the one-time setup link printed in the log), sign in, create an agent under Settings → People and copy its key. The code is MIT-licensed at &lt;a href="https://github.com/backmeupplz/agentboard" rel="noopener noreferrer"&gt;github.com/backmeupplz/agentboard&lt;/a&gt;, and there is a &lt;a href="https://agentboard.win/claude-code-kanban-board.html" rel="noopener noreferrer"&gt;guide to giving Claude Code a shared board&lt;/a&gt; if that is the agent you want to start with.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;h3&gt;
  
  
  How does an AI agent authenticate to agentboard?
&lt;/h3&gt;

&lt;p&gt;With a per-agent API key sent as &lt;code&gt;Authorization: Bearer ab_…&lt;/code&gt;. A human creates the agent user and the key in Settings → People.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can an agent create users, columns or keys?
&lt;/h3&gt;

&lt;p&gt;No. Those routes, and ticket deletion, answer &lt;code&gt;403 humans only&lt;/code&gt; for agent keys.&lt;/p&gt;

&lt;h3&gt;
  
  
  What happens to the old key when I rotate it?
&lt;/h3&gt;

&lt;p&gt;It returns &lt;code&gt;401&lt;/code&gt; immediately and any live event stream opened with it is closed.&lt;br&gt;
&lt;em&gt;Tested on 2026-11-06 against agentboard 0.1.13 (commit 22ca453) on Node 22.23.4, on a fresh local database.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>api</category>
      <category>security</category>
      <category>opensource</category>
    </item>
    <item>
      <title>How Veydrift's resource tokens work on Base: vMETAL, vCRYSTAL, vDEUT and the 28-day Rift lock</title>
      <dc:creator>borodutch.eth 💎🤲</dc:creator>
      <pubDate>Thu, 05 Nov 2026 11:00:15 +0000</pubDate>
      <link>https://dev.to/backmeupplz/how-veydrifts-resource-tokens-work-on-base-vmetal-vcrystal-vdeut-and-the-28-day-rift-lock-45c3</link>
      <guid>https://dev.to/backmeupplz/how-veydrifts-resource-tokens-work-on-base-vmetal-vcrystal-vdeut-and-the-28-day-rift-lock-45c3</guid>
      <description>&lt;p&gt;Veydrift's Metal, Crystal and Deuterium exist twice: as numbers on your planets inside the game, and as three ERC-20 tokens on Base called vMETAL, vCRYSTAL and vDEUT. One token is exactly one unit in the game, and the only way between the two is a building called the Rift Stabilizer. Putting tokens in is instant. Taking resources out takes 28 days, during which anyone can raid them. This post walks through how that works in the contracts, with the numbers I read from Base on November 5, 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the three tokens are
&lt;/h2&gt;

&lt;p&gt;Each token is an instance of one small contract, &lt;code&gt;VeydriftResource&lt;/code&gt;, which is about twenty lines of Solidity on top of OpenZeppelin's ERC-20. Its constructor mints 50,000,000,000 units once to a single holder, and there is no owner, no minter, no proxy and no upgrade path afterwards. It also overrides &lt;code&gt;decimals()&lt;/code&gt; to return 0, which is the whole trick behind "one token equals one unit": a balance of 500,000 vMETAL is 500,000 Metal, with no 18-decimal arithmetic in between.&lt;/p&gt;

&lt;p&gt;I called &lt;code&gt;totalSupply()&lt;/code&gt; on all three on Base today and each returned 50,000,000,000, and &lt;code&gt;decimals()&lt;/code&gt; returned 0. The addresses are listed on the &lt;a href="https://veydrift.com/docs/contracts" rel="noopener noreferrer"&gt;Contracts and Tokens page&lt;/a&gt; of the docs, and they are the redeployed set from October 21, 2026. An earlier set with 10 billion units and 6 decimals was retired on October 21, so if you find an older vMETAL address in a search result, it is not the one the game uses.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the supply sits
&lt;/h2&gt;

&lt;p&gt;Two addresses hold all of it. When I read the balances today, the game contract held 2,000,000,000 of each token and the treasury held 48,000,000,000 of each. That adds up to the 50 billion minted, so nothing has leaked anywhere else yet.&lt;/p&gt;

&lt;p&gt;The 2 billion in the game contract is the Rift's reserve. The contract keeps two counters, &lt;code&gt;totalInternalResources&lt;/code&gt; (what players hold on planets) and &lt;code&gt;lockedWithdrawalResources&lt;/code&gt; (what sits in a Rift extraction), and the sum of the two is the amount the reserve has to cover. Every time the game credits resources to a planet from a deposit, &lt;code&gt;_requireReserveCapacity&lt;/code&gt; checks that the contract's token balance still covers that sum.&lt;/p&gt;

&lt;p&gt;Here is what the counters said today, next to the reserve:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Token&lt;/th&gt;
&lt;th&gt;On planets&lt;/th&gt;
&lt;th&gt;In extractions&lt;/th&gt;
&lt;th&gt;Reserve&lt;/th&gt;
&lt;th&gt;Reserve ÷ liabilities&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;vMETAL&lt;/td&gt;
&lt;td&gt;77,038,061&lt;/td&gt;
&lt;td&gt;21,422,952&lt;/td&gt;
&lt;td&gt;2,000,000,000&lt;/td&gt;
&lt;td&gt;20.3&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;vCRYSTAL&lt;/td&gt;
&lt;td&gt;65,346,363&lt;/td&gt;
&lt;td&gt;7,942,502&lt;/td&gt;
&lt;td&gt;2,000,000,000&lt;/td&gt;
&lt;td&gt;27.3&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;vDEUT&lt;/td&gt;
&lt;td&gt;61,104,753&lt;/td&gt;
&lt;td&gt;9,138,544&lt;/td&gt;
&lt;td&gt;2,000,000,000&lt;/td&gt;
&lt;td&gt;28.5&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;So every unit of Metal, Crystal and Deuterium that exists in the game today is backed by a token in the reserve, and the reserve is twenty to twenty-eight times what the game currently owes. When players produce more than the reserve can cover, the treasury tops it up with a plain token transfer, and the docs say that is what the other 48 billion is for.&lt;/p&gt;

&lt;h2&gt;
  
  
  Putting tokens into the game
&lt;/h2&gt;

&lt;p&gt;Injecting is the easy direction. You need a Rift Stabilizer on the planet (it costs 8,000 Metal, 8,000 Crystal and 4,000 Deuterium, and needs Robotics Factory 4, Research Lab 2, Energy Technology 5 and Hyperspace Technology 1), and then it is two wallet transactions: an ERC-20 &lt;code&gt;approve&lt;/code&gt; for exactly the amount, followed by &lt;code&gt;depositMarketResource(planetId, resource, amount)&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;In &lt;code&gt;VeydriftPlanetManagementModule.sol&lt;/code&gt; that function checks you own the planet, checks the Stabilizer, pulls the tokens in with &lt;code&gt;transferFrom&lt;/code&gt;, and then compares the contract's balance before and after. If the balance did not rise by at least &lt;code&gt;amount&lt;/code&gt;, it reverts, so a token that takes a fee on transfer cannot trick the accounting. Then it adds the amount to the planet's resources in the same transaction. Injected resources are spent exactly like mined ones, and you can only inject into your own planets: to help a friend you send them tokens and they inject.&lt;/p&gt;

&lt;h2&gt;
  
  
  Taking resources out
&lt;/h2&gt;

&lt;p&gt;Extraction is deliberately the opposite: slow and public.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;startRiftExtraction(planetId, resource, amount)&lt;/code&gt; spends the amount from the planet and adds it to &lt;code&gt;_riftLockedResources&lt;/code&gt; for that planet. It records &lt;code&gt;unlocksAt = startedAt + RIFT_EXTRACTION_DELAY&lt;/code&gt;, and the constant in &lt;code&gt;VeydriftGameStorage.sol&lt;/code&gt; is &lt;code&gt;28 days&lt;/code&gt;. A planet can run one extraction per resource, and there is no function to cancel one.&lt;/li&gt;
&lt;li&gt;While it is locked, the planet is a legitimate target even when new-player protection, low-score protection or the bashing limit would normally block an attack. The attack module checks whether the lock is non-zero and, if so, lets the attack through but restricts the raid to the locked amount only. Your own alliance still cannot attack you, and moon attacks cannot reach a planet's lock.&lt;/li&gt;
&lt;li&gt;In &lt;code&gt;raidRiftExtraction&lt;/code&gt; a winning attacker takes the normal loot first and then fills spare cargo from the lock. There is no 50% ceiling on the lock the way there is on ordinary resources, so a fleet with enough cargo can take all of it. That function moves the looted amount from the locked counter to the in-game counter without touching the reserve, and it reverts unless the game contract itself calls it, so nobody can call it directly to wipe someone else's lock.&lt;/li&gt;
&lt;li&gt;After 28 days &lt;code&gt;finalizeRiftExtraction&lt;/code&gt; pays whatever survived to the planet owner's wallet as tokens. It does not need a Stabilizer, so you can finalize from any of your planets.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The pieces fit one idea: tokens can enter at any time, but leaving costs a month of visible exposure, so a planet cannot quietly drain its stockpile into a wallet the moment an enemy fleet is inbound. Everyone can see who is extracting, because the extractions show up in the Raid Finder under Rifters and in the contract's &lt;code&gt;RiftExtractionStarted&lt;/code&gt; events.&lt;/p&gt;

&lt;p&gt;One consequence is easy to check on chain. The reserve still holds exactly the 2,000,000,000 it started with, so no extraction has been paid out yet, which fits the clock: the tokens were deployed on October 21, and the earliest any extraction can finalize is 28 days after it starts.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you can verify yourself
&lt;/h2&gt;

&lt;p&gt;Three properties come out of this layout, and each one can be checked in the contracts rather than taken from me.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The tokens have no owner.&lt;/strong&gt; Nobody can mint more vMETAL or freeze a balance, so the supply numbers above are final.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backing is checked by code.&lt;/strong&gt; The game cannot credit resources it cannot cover, because the same function that credits them reads the token balance.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Extraction risk is part of the rules.&lt;/strong&gt; The 28-day lock and the raid rule live in the same contracts as combat, so there is no off-chain setting that changes them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you want to read the code yourself, start with &lt;code&gt;VeydriftResource.sol&lt;/code&gt; (twenty lines), then &lt;code&gt;VeydriftRiftModule.sol&lt;/code&gt; for the extraction lifecycle, and &lt;code&gt;VeydriftResourceReserves.sol&lt;/code&gt; for the reserve requirement. The player-facing version of all this is the Rift section of the &lt;a href="https://veydrift.com/docs" rel="noopener noreferrer"&gt;Veydrift docs&lt;/a&gt;, and the live activity is on &lt;a href="https://stats.veydrift.com" rel="noopener noreferrer"&gt;stats.veydrift.com&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions people ask
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Can I trade vMETAL, vCRYSTAL and vDEUT?&lt;/strong&gt; They are ordinary ERC-20 tokens, so you can hold them, send them to another player or inject them into any of your planets. The docs list the three contract addresses, and say the treasury adds resource pools against $VEYDRIFT after the token auction.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What happens to resources if my planet is raided during an extraction?&lt;/strong&gt; The raid takes its normal loot first and then draws from the locked amount. Whatever is left when the 28 days end is paid to your wallet.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I cancel an extraction?&lt;/strong&gt; No. Once started, the amount stays locked until it is finalized.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why 0 decimals?&lt;/strong&gt; So that a token balance and an in-game balance are the same number, with no rounding between them.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Checked against the Veydrift repository at commit 4e54b54a (November 4, 2026) and the game contract on Base on November 5, 2026. Nothing here is investment advice: it describes how the contracts behave.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://blog.veydrift.com/veydrift-resource-tokens-base-rift/" rel="noopener noreferrer"&gt;the Veydrift blog&lt;/a&gt;. Veydrift is an &lt;a href="https://veydrift.com" rel="noopener noreferrer"&gt;onchain space strategy game on Base&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>web3</category>
      <category>solidity</category>
      <category>gamedev</category>
      <category>blockchain</category>
    </item>
  </channel>
</rss>
