<?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:  AgentHardWay</title>
    <description>The latest articles on DEV Community by  AgentHardWay (@agenthardway).</description>
    <link>https://dev.to/agenthardway</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%2F4258922%2F658e5350-7682-403e-bb29-c773b7dc128f.png</url>
      <title>DEV Community:  AgentHardWay</title>
      <link>https://dev.to/agenthardway</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/agenthardway"/>
    <language>en</language>
    <item>
      <title>Hermes Agent vs Claude Code, split by evidence</title>
      <dc:creator> AgentHardWay</dc:creator>
      <pubDate>Sun, 22 Nov 2026 15:18:07 +0000</pubDate>
      <link>https://dev.to/agenthardway/hermes-agent-vs-claude-code-split-by-evidence-50mn</link>
      <guid>https://dev.to/agenthardway/hermes-agent-vs-claude-code-split-by-evidence-50mn</guid>
      <description>&lt;p&gt;Searches for "hermes agent vs claude code" want a verdict. What they can get here is one measured half, one documented half, and an explicit line between them — because a comparison that mixes the two is how the genre usually goes wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule: two evidence levels, never one column
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Measured&lt;/strong&gt; means we ran it on this machine and can rerun it. &lt;strong&gt;Documented&lt;/strong&gt; means the vendor's own page says so on a stated date. They are not interchangeable, and the one thing this post will not do is average them into a single score.&lt;/p&gt;

&lt;p&gt;The Claude Code column was read on 2026-11-22 from &lt;code&gt;docs.claude.com&lt;/code&gt;. It was not installed, not logged in, and no task was run on it.&lt;/p&gt;

&lt;p&gt;If you are still at the wider question — which agents exist at all, and how each is paid for — that comparison is &lt;a href="https://agenthardway.com/cost/hermes-agent-alternatives/" rel="noopener noreferrer"&gt;one page over&lt;/a&gt;, with the same split.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install, and what lands on disk
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Hermes Agent · measured&lt;/th&gt;
&lt;th&gt;Claude Code · documented&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Install&lt;/td&gt;
&lt;td&gt;&lt;code&gt;uv tool install&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;native installer (&lt;code&gt;claude.ai/install.sh&lt;/code&gt;), WinGet, Homebrew casks, apt/dnf/apk&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;What you get&lt;/td&gt;
&lt;td&gt;three executables and a 116 MB directory&lt;/td&gt;
&lt;td&gt;a CLI plus IDE extensions, a desktop app and a web surface&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Updates&lt;/td&gt;
&lt;td&gt;you re-run the upgrade path; we back up before it&lt;/td&gt;
&lt;td&gt;native installs update in the background; Homebrew and WinGet installs do not&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Windows note&lt;/td&gt;
&lt;td&gt;installs natively; we also run the WSL backend&lt;/td&gt;
&lt;td&gt;Git for Windows recommended on native Windows so the Bash tool has a shell; WSL needs it not&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The Claude Code row is more than we can verify: "native installs update in the background" is a behaviour we have not watched, and it is also the kind of behaviour that changes what "which version am I on" means.&lt;/p&gt;

&lt;h2&gt;
  
  
  How each one charges you
&lt;/h2&gt;

&lt;p&gt;This is the structural difference, and it is not a quality difference:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hermes Agent&lt;/strong&gt; is a client for a model API. You pay your provider per token, and every unit price is on a page you can read. That makes cost a multiplication you can do before you spend.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code&lt;/strong&gt; is billed through a Claude subscription or an Anthropic Console account. A plan caps the bill; it does not expose a per-token number to multiply.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Which is better depends on whether your usage is bursty (a cap helps) or steady and measurable (a rate helps). And on 2026-11-22 the second half of that sentence could not be checked: &lt;code&gt;docs.claude.com/en/docs/about-claude/pricing&lt;/code&gt; answered &lt;strong&gt;"App unavailable in region"&lt;/strong&gt;. So this post contains no Claude Code price at all — not because it is unflattering, but because we could not read it.&lt;/p&gt;

&lt;h2&gt;
  
  
  One task, two configurations, our numbers
&lt;/h2&gt;

&lt;p&gt;Fixed workload: 200,000 input tokens, 90% cache hit, 20,000 output tokens.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;deepseek-flash, off-peak     $0.0155
deepseek-v4-pro, peak        $0.1135      7.3x
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The 7.3× is cheapest configuration versus priciest on the same task. The peak-versus-off-peak multiplier is &lt;strong&gt;2×&lt;/strong&gt; — a different number that gets quoted as 7.3× often enough to be worth repeating.&lt;/p&gt;

&lt;p&gt;Those rates come from a price file transcribed from the provider's own page with a capture date. The cache-hit rate is ours, not a vendor claim: &lt;strong&gt;99.1% over 1,076 calls in a 30-day window&lt;/strong&gt;, which is what makes the cheap configuration cheap in practice rather than in theory.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three questions that decide it
&lt;/h2&gt;

&lt;p&gt;Not a score — the three forks where the two products actually diverge:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Do you need to know the cost before you spend, or only after?&lt;/strong&gt; Per-token pricing answers "what will this task cost" before the run. A subscription answers "what did this month cost" after it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Is your work Windows-native?&lt;/strong&gt; Both document Windows. Hermes' install footprint is one directory you can delete; Claude Code's docs recommend an extra dependency (Git for Windows) for its shell tool on native Windows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Do you want the model to be your choice?&lt;/strong&gt; Hermes is provider-agnostic by construction — the bill is the provider's. Claude Code's docs say its terminal and IDE surfaces also accept third-party providers, which narrows this difference but does not remove it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What this post will not tell you
&lt;/h2&gt;

&lt;p&gt;Which one writes better code, handles a large repo better, or recovers better from a failed edit. All three need the same task run in both, and we have not installed Claude Code. If a page gives you those without running both, it is reporting someone else's numbers with its own name on top.&lt;/p&gt;

&lt;p&gt;We also could not read their pricing page from this region on 2026-11-22, so any cost comparison that includes a subscription figure is out of scope here by construction, not by choice.&lt;/p&gt;

&lt;h2&gt;
  
  
  How you know it worked
&lt;/h2&gt;

&lt;p&gt;For the reader this is a protocol, not a claim: run the same task in both, then take tokens from &lt;strong&gt;your own&lt;/strong&gt; provider billing page and multiply by the rate you pay. The check that catches self-deception is to run the same task twice in the same agent on different days — if your per-task cost moves more than the rate list allows, you are measuring your prompt, not the vendor.&lt;/p&gt;

&lt;p&gt;For the subscription side, the equivalent check is the cap: log used-vs-allowed before and after a fixed task, and see whether the task moved it by what you expected. We have not run that check on Claude Code.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to undo it
&lt;/h2&gt;

&lt;p&gt;Pieces, in the order that does not lose data:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Keep the state, remove the program. On this stack the install directory is disposable; &lt;code&gt;%LOCALAPPDATA%\hermes&lt;/code&gt; (Windows) or &lt;code&gt;~/.hermes&lt;/code&gt; (WSL) is not — memory, skills and per-profile session databases live there.&lt;/li&gt;
&lt;li&gt;Copy the state directory before touching anything, and only delete the copy once the replacement is working.&lt;/li&gt;
&lt;li&gt;Revert the model configuration separately from the install: a client pointed at a dead endpoint looks exactly like a broken agent.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;We have run backup and upgrade on this stack. &lt;strong&gt;We have not run an uninstall, and we have not tested migration into Claude Code.&lt;/strong&gt; Both are statements about our evidence, not about the products.&lt;/p&gt;

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

&lt;p&gt;Everything on the Claude Code side of the table: install, login, tool use, cost, and the background-update behaviour its docs describe. Also untested: cross-tool migration in either direction, and any claim about output quality on either side.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://agenthardway.com/cost/hermes-agent-vs-claude-code/" rel="noopener noreferrer"&gt;agenthardway.com/cost/hermes-agent-vs-claude-code&lt;/a&gt;. The split-by-evidence rule this post follows is written up at &lt;a href="https://agenthardway.com/method/" rel="noopener noreferrer"&gt;agenthardway.com/method&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claudecode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Agent alternatives, split by who measured what</title>
      <dc:creator> AgentHardWay</dc:creator>
      <pubDate>Sun, 22 Nov 2026 15:18:05 +0000</pubDate>
      <link>https://dev.to/agenthardway/agent-alternatives-split-by-who-measured-what-39g2</link>
      <guid>https://dev.to/agenthardway/agent-alternatives-split-by-who-measured-what-39g2</guid>
      <description>&lt;p&gt;People search "hermes agent alternatives" when they are deciding whether to switch. The honest answer has two halves with different reliability, so the table below is split by who measured what: &lt;strong&gt;one column we ran, three columns we only read&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;That distinction is the whole point. A comparison that quietly mixes a number you produced with a number you copied off a vendor page is how this genre usually goes wrong — and the reader cannot tell which line is which.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule this page follows
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;A number we did not produce is labelled as documentation, not as a result.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The three competitor columns were read on 2026-11-22 from &lt;code&gt;docs.claude.com&lt;/code&gt;, &lt;code&gt;opencode.ai/docs&lt;/code&gt; and &lt;code&gt;github.com/openai/codex&lt;/code&gt;. &lt;strong&gt;None of them was installed, authenticated or run here.&lt;/strong&gt; If a sentence about those tools is wrong, it is because their docs changed, not because we tested it.&lt;/p&gt;

&lt;h2&gt;
  
  
  How you install and how you pay
&lt;/h2&gt;

&lt;p&gt;These are the two questions that decide most evaluations, and they are also the two the vendors document precisely.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Agent · evidence&lt;/th&gt;
&lt;th&gt;Install&lt;/th&gt;
&lt;th&gt;How you pay&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Hermes Agent&lt;/strong&gt; · measured&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;uv tool&lt;/code&gt; on Windows; three executables, 116 MB&lt;/td&gt;
&lt;td&gt;Per token, on your provider's API (DeepSeek here)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Claude Code&lt;/strong&gt; · docs&lt;/td&gt;
&lt;td&gt;native installer, Homebrew, WinGet, apt&lt;/td&gt;
&lt;td&gt;Claude subscription or Console account; third-party providers allowed&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;OpenCode&lt;/strong&gt; · docs&lt;/td&gt;
&lt;td&gt;npm, brew, bun, mise — no Windows package managers&lt;/td&gt;
&lt;td&gt;Bring your own provider, or a $10/month plan&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Codex CLI&lt;/strong&gt; · docs&lt;/td&gt;
&lt;td&gt;native installer, npm, brew cask&lt;/td&gt;
&lt;td&gt;ChatGPT plan, or an API key&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The exact commands, as each vendor writes them on 2026-11-22 (only the first was run here):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Hermes      uv tool install &amp;lt;package&amp;gt;
Claude Code curl -fsSL https://claude.ai/install.sh | bash
OpenCode    npm install -g @opencode/cli
Codex CLI   curl -fsSL https://chatgpt.com/codex/install.sh | sh
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Read the first column as the evidence tag: &lt;code&gt;measured&lt;/code&gt; is our run, &lt;code&gt;docs&lt;/code&gt; is the vendor page fetched 2026-11-22 and never executed here.&lt;/p&gt;

&lt;p&gt;Two structural differences fall out of that table, and neither is a quality judgement:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Subscription versus per-token.&lt;/strong&gt; Three of the four can be paid for as a plan; Hermes cannot, because it is a client for a model API. A plan caps the bill; a price list makes the bill computable. Which is better depends on whether your usage is bursty.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Which one you can install on Windows at all.&lt;/strong&gt; OpenCode's own docs say Windows package managers are not supported — a fact about deployment, not about the product.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What a task actually costs, in the one column we can compute
&lt;/h2&gt;

&lt;p&gt;For Hermes on DeepSeek the marginal cost is a multiplication: tokens in, cache-hit rate, tokens out, times the rate on the provider page. One fixed workload — 200,000 input tokens, 90% cache hit, 20,000 output tokens — measured on this machine:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;deepseek-flash, off-peak     $0.0155
deepseek-v4-pro, peak        $0.1135      7.3x
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The off-peak multiplier is 2×, not 7.3×. The 7.3× is cheapest configuration versus priciest &lt;strong&gt;on the same task&lt;/strong&gt;, which is a different claim and the one people usually mistype.&lt;/p&gt;

&lt;p&gt;For the other three, this post can tell you the billing &lt;em&gt;model&lt;/em&gt; and not the &lt;em&gt;number&lt;/em&gt;, because the number needs your token counts and your plan. Anyone publishing a per-task dollar figure for a subscription product without those is inventing the denominator.&lt;/p&gt;

&lt;p&gt;The most-searched pairing — Hermes Agent against Claude Code — is split the same way on &lt;a href="https://agenthardway.com/cost/hermes-agent-vs-claude-code/" rel="noopener noreferrer"&gt;its own page&lt;/a&gt;, with their pricing page quoted as unreadable on the day we tried.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you give up if you switch away
&lt;/h2&gt;

&lt;p&gt;Not "features" — the four things we can point at with a file or a command:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Thing&lt;/th&gt;
&lt;th&gt;Where it is&lt;/th&gt;
&lt;th&gt;After a switch&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Recomputable per-token cost&lt;/td&gt;
&lt;td&gt;price file + your provider bill&lt;/td&gt;
&lt;td&gt;Subscription products do not expose unit prices; the multiplication disappears&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Three separate stores&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;%LOCALAPPDATA%\hermes&lt;/code&gt;, &lt;code&gt;~/.hermes&lt;/code&gt;, per-profile dirs&lt;/td&gt;
&lt;td&gt;Each product has its own layout; migration untested&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Error index by verbatim text&lt;/td&gt;
&lt;td&gt;&lt;a href="https://agenthardway.com/symptoms/" rel="noopener noreferrer"&gt;the symptom index&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Rebuilding it is the expensive part, not installing&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gates that decide "done"&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;check&lt;/code&gt; / &lt;code&gt;postbuild&lt;/code&gt; / &lt;code&gt;deploy&lt;/code&gt; scripts&lt;/td&gt;
&lt;td&gt;Written against this stack's artifacts; they do not travel&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That list is why "just try the other one for a week" is cheap for a hobby project and expensive for a working setup.&lt;/p&gt;

&lt;h2&gt;
  
  
  How you know it worked
&lt;/h2&gt;

&lt;p&gt;If you are evaluating rather than reading, the only number worth trusting is the one you produce. Run the same task in two agents, take tokens from &lt;strong&gt;your own&lt;/strong&gt; provider billing page, and multiply by the rate you actually pay. Three fields decide it: input tokens, cache-hit tokens, output tokens.&lt;/p&gt;

&lt;p&gt;The check that catches self-deception: run the same task twice in the same agent on different days. If your cost per task moves more than the price list says it should, you are measuring your prompt, not the vendor.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to undo it
&lt;/h2&gt;

&lt;p&gt;Program and state are separate on this stack: the install directory is disposable, the state is not. On Windows it is &lt;code&gt;%LOCALAPPDATA%\hermes&lt;/code&gt;; under WSL it is &lt;code&gt;~/.hermes&lt;/code&gt;; named profiles keep their own memory and session databases underneath. Removing the executable leaves all of it in place.&lt;/p&gt;

&lt;p&gt;We have run the backup and the upgrade path, and we have not run an uninstall. So treat "uninstall is clean" as unverified — the safe order is: copy the state directory somewhere first, remove the program, then delete the copy once the replacement works.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we did not run
&lt;/h2&gt;

&lt;p&gt;The whole competitor half. No install, no login, no task, no cost measurement for Claude Code, OpenCode or Codex CLI. Their columns are what their documentation says on 2026-11-22, and nothing more.&lt;/p&gt;

&lt;p&gt;We also did not test cross-tool migration in either direction. If you need that answered before switching, this post cannot answer it, and neither can a table.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://agenthardway.com/cost/hermes-agent-alternatives/" rel="noopener noreferrer"&gt;agenthardway.com/cost/hermes-agent-alternatives&lt;/a&gt; — the numbers above are regenerated from a price file on every build, and the method behind them is documented at &lt;a href="https://agenthardway.com/method/" rel="noopener noreferrer"&gt;agenthardway.com/method&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
  </channel>
</rss>
