TokenWatch

Chrome extension (MV3, plain JS, no build) that shows your AI subscription usage limits and API credits in one popup, using the sessions you're already logged in with. No API keys, no analytics, no remote code.

Install (load unpacked)

1. Open chrome://extensions.
2. Turn on Developer mode (top right).
3. Click Load unpacked and pick this folder (the one with manifest.json).
4. Pin TokenWatch from the puzzle-piece menu and click it.

After editing any file, click the ↻ reload icon on the TokenWatch card in chrome://extensions.

What it reads (captured 2026-10-16/17)

Service	Requests	Auth
Claude	GET /api/organizations → picks the org with chat capability; plan from claude_pro / claude_max GET /api/organizations/{uuid}/usage → limits[] (kind, percent, resets_at)	cookies only
ChatGPT	GET /api/auth/session → accessToken (kept in memory only, max 10 min) GET /backend-api/wham/usage → rate_limit.primary_window (5-hour) / secondary_window (weekly), labeled by limit_window_seconds: used_percent, reset_at (unix seconds)	Authorization: Bearer <accessToken>
Gemini	GET /app (HTML) → XSRF token SNlM0e (kept in memory only, max 10 min) POST /_/BardChatUi/data/batchexecute?rpcids=jSf9Qc&rt=c, form body f.req=[[["jSf9Qc","[]",null,"generic"]]]&at=<token> → [2, [[?, fractionUsed, kind (1 = current, 2 = weekly), [[resetUnixSeconds, nanos]]], …], false]	cookies + at token
Claude API	GET /api/organizations on platform.claude.com → picks the org with api capability GET /api/organizations/{uuid}/current_spend → amount (cents), resets_at GET /api/organizations/{uuid}/spend_limits → spend_limits[].limit_usd (cents; despite the name), enforced_limit_usd (tier cap) GET /api/organizations/{uuid}/prepaid/credits → amount (cents)	cookies only
OpenAI API	Runs inside an open platform.openai.com tab: reads localStorage["oai-auth:onboard"].accessToken, POST api.openai.com/v1/dashboard/onboarding/login → user.session.sensitive_id (sess-…) GET api.openai.com/v1/dashboard/billing/credit_grants → total_granted, total_used, total_available (dollars)	page tokens; never leave the tab

Claude and ChatGPT show 5-hour limit and Weekly limit. Claude Max also shows per-model weekly limits, such as Weekly limit (Fable), and its tier, such as Max 5x. ChatGPT Pro has no 5-hour limit, so it shows only Weekly limit. Gemini shows Current limit and Weekly limit, which are the two bars on gemini.google.com/usage; Gemini doesn't say how long its current window is. Gemini's plan isn't shown because it comes from a request that hasn't been captured yet. ChatGPT's numbers are the "Plan limits" on chatgpt.com/settings/usage?tab=overview, which ChatGPT says are shared across Codex, Work, Workspace Agents and ChatGPT for Excel. Regular chat conversations are not included.

Claude API shows Monthly spend against your own spend limit (or the tier cap when none is set), which is the bar on platform.claude.com/dashboard, and Credit balance, which is the total on Settings → Billing, promotional credits included. OpenAI API shows Credit balance (the "Credit remaining" on platform.openai.com/home) with a bar for how much of your granted credit is used. It only refreshes while a platform.openai.com tab is open, because the dashboard login is kept in that page rather than in cookies. For that reason it's off by default; turn it on in Options.

Permissions

• storage: settings and parsed results (chrome.storage.local); the Debug raw JSON goes in chrome.storage.session (memory, cleared when Chrome closes).
• alarms: background refresh (default 15 min, minimum 5).
• scripting: the fallback. When a background fetch fails (cookies not sent, Cloudflare, missing page headers), the same request is run inside an already-open tab of that site (MAIN world). The extension never opens tabs on its own. It finds the tab through chrome.tabs.query with a URL filter, which the host permissions already allow, so the tabs permission isn't needed.
• Host permissions: https://claude.ai/*, https://chatgpt.com/*, https://gemini.google.com/*, https://platform.claude.com/*, https://platform.openai.com/* only.

Choose which services appear, and their order in the popup, in Options → Services to show. Options → Toolbar badge shows the highest % used across all services by default. Tick providers, metrics (such as Weekly limit, which then covers every ticked provider's weekly limit), or both to narrow it; the badge shows the highest of what's ticked.

Politeness

Fetches happen when the popup opens (the cache is reused if it's under 1 minute old), when you click Refresh, and on the alarm. Any non-OK result, including HTTP 429, backs off automatically: 5 → 10 → 20 … minutes, capped at 2 h. The manual Refresh button ignores backoff.

Tests

node --test tests/parsers.test.mjs: parser tests built from the captured responses.

Troubleshooting

• "Not logged in": log in on the site (the button opens it), then press Refresh. If you are logged in, Chrome may not have sent your cookies to the background request. Keep a tab of the site open and press Refresh so the in-tab fallback runs.
• 401 / 403: the session expired, or the org changed (Claude re-discovers its org on the next fetch after a 403/404). For ChatGPT, a 401 gets one automatic retry with a fresh token. If it keeps failing, reload the site tab and refresh.
• Cloudflare challenge ("Blocked by a Cloudflare check"): open or reload the site in a tab, complete any check, then Refresh. The in-tab fallback reuses the cleared session.
• "Couldn't parse" / endpoint changed:
  1. Tick Debug in the popup to see the raw JSON from the last fetch.
  2. Open the site's usage page with DevTools → Network → Fetch/XHR and find the new request (Claude: Settings → Usage, look for usage; ChatGPT: Settings → Usage, look for wham/usage; Gemini: Settings → Usage limits, look for batchexecute with rpcids=jSf9Qc).
  3. Edit only that service's adapter: adapters/claude.js, adapters/chatgpt.js, adapters/gemini.js, adapters/claude-api.js or adapters/openai-api.js. The URLs are at the top of fetchUsage(); the mapping is in parseUsage(). Both must return the same normalized shape.
  4. Update the sample in tests/parsers.test.mjs, run the tests, and reload the extension.