DEV Community

정의영
정의영

Posted on

I built a local-first tab manager and gave it an MCP server so Claude can organize my tabs

My browser is a war crime

I'm a designer and developer, and a hopeless tab hoarder. On a normal day I've got 70+ tabs open across a couple of windows, every favicon crushed into an unreadable sliver, and my "system" for finding anything is to just… open another tab. Bookmarks? A write-only graveyard I never revisit.

I bounced between tab managers for a while. Toby was the one that stuck — until it capped the free plan at 60 saved tabs and started asking for an account. For a tab hoarder, a 60-tab cap is like a notes app that locks after 60 words. That was the nudge.

So I did the reasonable thing and spent way too long building my own. It's called Mos Tab. This is the build story — the parts I'm proud of, the part that nearly broke me (multiple windows, I'm looking at you), and the weirdly fun part at the end where I let Claude reorganize my tabs.

What I actually wanted

Three non-negotiables:

  1. Local-first. No account required. Data lives in the browser, works offline. I was tired of "sign in to see your own bookmarks."
  2. No limits on the free stuff. Saving a tab locally costs me nothing to run — so there's no reason to cap it.
  3. A calm dashboard, not a list. I stare at my new tab all day; it should feel like a desk, not a void.

The shape of it

Mos Tab overrides chrome://newtab (MV3). The data model is a boring normalized tree — Space → Collection → Section → Tab — records plus ID arrays for ordering. It's a single Zustand store with a persist middleware writing to chrome.storage.local. Your currently-open tabs show in a live side panel (chrome.tabs); one click saves them into a collection, drag to organize, and a cmdk palette searches everything you've saved.

Stack, for the curious: React 18 + TypeScript + Vite + @crxjs/vite-plugin, Tailwind, dnd-kit, zod, i18next (5 languages).

Since the new tab is prime real estate, I threw in the stuff I open a tab for anyway — calendar, pomodoro, to-dos, sticky notes. All free.

So far, so normal. Then I wanted sync.

The hard part: sync (or, how multiple windows tried to kill me)

Local-first is lovely until you have two machines. I wanted optional cloud sync (the paid tier — more on money later) without turning this into a database migration.

The backend is deliberately dumb: a Cloudflare Worker + R2, one JSON blob per user (state.json). No per-row DB. To save, you PUT the blob back with an ETag conditional request (If-Match). On a 412 (someone wrote first), you pull, 3-way merge on the client, and push again.

The 3-way merge (base + local + remote) handles the real cases: both sides added a tab → keep both; something in base vanished on one side → it was deleted, respect it; field conflicts → local wins. Sticky-note text conflicts resolve to a symmetric content-addressed copy, so two machines converge instead of endlessly duplicating.

The part that actually broke me wasn't cross-device — it was multiple windows on the same machine. Every window wanted to push, each invalidating the others' ETag → 412 → merge → push → 412 → forever. A beautiful infinite loop.

The fix: only the focused window pushes. Every window can pull (reads are idempotent), but writes are gated to the focused one, with a flush-on-blur so the last edit isn't lost when you switch away. Add adaptive polling backoff (15s active → 2min idle) and an "away mode" that stops polling after 20 min of no activity, and it stops hammering the Worker.

The fun part: teaching Claude to use my tabs (MCP)

Here's the part I'm actually excited about. I gave Mos Tab a remote MCP server so you can connect Claude or ChatGPT to it. Once connected, the assistant can search and summarize everything you've saved (tabs, notes, to-dos) and — the fun bit — save new things it finds into a brand-new space, add notes, or make a public share link. All from the chat.

Concrete example. I told Claude:

"Find the top-rated budget strollers on Amazon, save them to my tabs, and make a share link."

A few seconds later there was a new "Stroller shortlist" space on my dashboard, organized into categories, and a public link I sent straight to my wife — she opened it in her browser, nothing to install.

Two design decisions I'm happy with:

  • It's add-only. The assistant can create spaces, notes, and shares — but it can't touch or reorganize your existing spaces. No "the AI nuked my setup" risk. That made me trust it way more.
  • The implementation is boring on purpose. Streamable HTTP JSON-RPC on the same plain-fetch Worker — no Durable Objects, no McpAgent. Auth is a long-lived scoped token you mint in settings (JWTs expire, which is bad for a persistent connection). Reads filter the same per-user blob in memory; writes are read-modify-write with If-Match / 412 retry — so an AI-created space merges through the exact same 3-way sync as a manual edit.

That last bit is the thing I like most: AI writes go through the same door as everything else. No special path, no separate store.

Money, honestly

Because someone always asks: the whole local app is free — unlimited tabs, spaces, all the widgets, import/export. Local storage costs me nothing to run, so there's no reason to cap it. The paid tier is only the stuff that actually hits a server: multi-device sync, public sharing, and the AI/MCP integration. I can't give the server stuff away free — kind of wish I could. That's the whole pricing philosophy, and it keeps me honest.

What I learned

  • "Local-first" is a great default and a real constraint. Deciding exactly what syncs vs. what stays on-device (your data and notes sync; theme and zoom are per-device) took more thought than the merge algorithm.
  • The boring backend was the right call. One blob + conditional PUT scales further than I expected — the 5MB blob cap only bites around 15–30k saved tabs, and by then I'd happily move to per-entity storage (D1).
  • MCP is more fun than it has any right to be. Giving an assistant a small, safe, add-only surface into your own data feels like the actual point of these protocols.

Try it

Live on the Chrome Web Store (Chrome + Chromium — Edge, Brave, Arc, etc.):

I'm a solo dev, so I read everything — feedback on the sync merge model or the MCP design especially welcome.

Top comments (1)

Collapse
 
deanlee profile image
Dean Lee

The add-only boundary is the part I would keep even if the sync layer changes. Agents are easier to trust when the worst case is a messy new space, not damage to the existing one. Did you consider making share-link creation a separate permission from saving tabs?