Stop watching your AI relearn the same task.
The open-source alternative to CoWork — 90% fewer tokens on repeat browser tasks.
If AgentLimb is useful to you, please ⭐ star this repo — it helps others discover the project and keeps us motivated!
English | 中文 | 日本語 | 한국어 | Español | Français | Deutsch | Português | Русский | العربية | Italiano | हिन्दी
AgentLimb is a browser extension — now live on the Chrome Web Store — that lets any AI terminal drive your browser with precision. It works with every Chromium-based browser (Chrome, Edge, Brave, Vivaldi, Chromium, …) and every major AI terminal — Claude Code, Codex, Cursor, Trae, Windsurf, Tencent WorkBuddy, Alibaba Qwen Office, ByteDance TRAE Work, Cola, Kimi Work, or any local model. Install the extension, copy one prompt, paste it to your AI, and it auto-configures in 10 seconds.
No headless browsers. No re-login. No invasive agents. Your real Chrome, your real cookies, your real sessions — plus muscle memory that makes repeat tasks dramatically cheaper every time you run them.
After the one-time local runtime install, copy a single prompt and paste it to any AI tool. No API keys or cloud account. If your AI can run commands, it can use AgentLimb.
The first time your AI hits a site, it explores the DOM and learns selectors / workflow. AgentLimb writes that knowledge to ~/Desktop/AgentLimb-muscle/<domain>.json. Every future run on the same site skips the exploration and reuses what was learned.
Real regression data from a Reddit posting task (low-parameter Codex, 2026-04-18):
| Cold start (first explore) | Hot start (muscle recall) | Savings | |
|---|---|---|---|
page_snapshot calls |
3 | 0 | 100% |
| Total tool calls | 23 | 10 | 56.5% |
| Estimated tokens | ~12,250 | ~1,750 | ↓ 85.7% |
| Wall-clock time | 8–20 min | 30 s–2 min | ↓ ~80–95% |
The more you reuse a site, the cheaper and faster it gets.
AgentLimb drives the browser through Chrome Debugger Protocol. The AI receives a structured semantic list of interactive elements — not screenshots. Clicks hit the right node, form fills use native APIs, navigations return the new URL immediately.
Silence no longer equals success. The AI explicitly declares task_plan → task_step_done → task_complete / task_fail. Timeouts and bridge drops are captured as real failures, not false positives. The side panel renders the live step list in real time.
Bridge runs on 127.0.0.1:7791. No analytics, no tracking, no cloud. Muscle knowledge is plain JSON on your desktop — you can read, diff, share, or delete it at any time.
One AI command, every Chrome profile executes it. Whether you have two Twitter accounts, three company Google accounts, or a dozen test profiles — AgentLimb drives them all in a single task.
- Explicit identity — each side panel shows "This panel: Profile-xxxxxx"; the Bridge tracks which profile sent which result
- Suspend / auto-suspend — close a panel (or click Suspend) to opt that profile out; the rest continue without interruption
- task_* fan-out —
task_plan,task_step_done,task_complete,task_failbroadcast to every active profile; all side panels stay in sync - Window locking —
navigatetargets the correct Chrome window automatically, even when the same profile has multiple windows open - Target routing — point a tool call at a specific profile by label for surgical precision
Your AI Terminal (Claude Code / Codex / Cursor / TRAE Work / WorkBuddy / Qwen Office / Kimi Work / local model)
↕ HTTP + SSE (16 standardized tools, auto-discoverable via /docs endpoints)
AgentLimb Bridge (local Node.js · 127.0.0.1:7791)
↕ chrome.runtime message passing
AgentLimb Extension (Chrome MV3 · side panel UI · task/muscle/log tabs)
↕ Chrome Debugger Protocol
Your Browser (logged in, with cookies, your real sessions)
↓ knowledge persisted
~/Desktop/AgentLimb-muscle/<domain>.json (durable, human-readable)
AgentLimb is agent-agnostic by design: if your AI can send an HTTP request or run a local command, it can drive the browser. It works on all platforms (Windows & macOS) with every Chromium-based browser.
International terminals — Claude Code, Codex, Cursor, Trae, Windsurf, and any local model.
Chinese agent platforms — fully supported:
| Platform | Highlights | |
|---|---|---|
| TOP 1 | Tencent WorkBuddy（腾讯 WorkBuddy） | Tencent's AI productivity expert with coding, data analysis, and office automation — pairs perfectly with AgentLimb for end-to-end web control |
| TOP 2 | Alibaba Qwen Office（阿里千问办公） | Alibaba's Qwen-powered workspace agent; give it a prompt and let it operate the web through your real browser sessions |
| TOP 3 | ByteDance TRAE Work（字节 TRAE Work） | ByteDance's AI-native work terminal; connect it to AgentLimb to turn chat into precise browser actions |
| TOP 4 | Cola | Lightweight Chinese agent client with strong terminal integration — one onboarding prompt and it's ready |
| TOP 5 | Kimi Work | Moonshot's Kimi Work agent handles long-horizon web tasks; muscle memory makes its repeat runs dramatically cheaper |
Universal web-control for every agent, on every Chromium browser. That's the goal.
Every GitHub Release ships two packages — make sure you grab the right one:
| Package | What it is | Who installs it |
|---|---|---|
agentlimb-chrome-v0.2.2-b8.zip |
Browser extension | You — install it into your Chromium browser |
agentlimb-runtime-v0.2.1-b7.zip |
Local Runtime (Bridge + CLI) | Your AI agent — downloaded automatically after your approval |
You normally never need to touch the Runtime zip manually. The onboarding prompt lets your AI download, verify (SHA-256), and install it for you.
- Chrome Web Store (recommended): Install AgentLimb. Works in Chrome, Edge, Brave, Vivaldi, Chromium, and any other Chromium-based browser.
- Manual install: download
agentlimb-chrome-v0.2.2-b8.zipfrom the Releases page, unzip it, then load the extracted folder on your browser's extensions page with Developer Mode enabled.
Open the AgentLimb side panel and click "Copy Onboard Prompt". Node.js 18 or later is required on your machine.
Any agent that can run local commands follows the same protocol: it asks for your explicit approval, then downloads the small Runtime from the GitHub Release, verifies its SHA-256 checksum, installs the local Bridge and Native Messaging host, and runs a health check. It does not scan browser profiles, folders, or extension sources.
What happens behind the scenes, per platform:
- macOS — the agent extracts
agentlimb-runtime-v0.2.1-b7.zip, then runsscripts/install.sh --extension-id <your-extension-id>. If AgentLimb is already installed, the agent simply starts the existing Runtime vialaunchctl— no re-download needed. - Windows — the agent extracts the same zip, then runs
powershell -ExecutionPolicy Bypass -File scripts\install.ps1 -ExtensionId <your-extension-id>.
Both installers register the Native Messaging host, install the agentlimb CLI, and start the local Bridge on 127.0.0.1:7791. A browser extension alone cannot do any of this — that's why the Runtime setup is deliberately handled by a terminal-capable agent.
Once the health check passes, the side panel shows ONLINE and your AI can drive every Chromium browser window through the local Bridge, discovering tool schemas on demand.
16 standardized tools spanning five categories: observing browser state, navigating and interacting with page elements, reading and writing muscle memory, declaring task lifecycle events, and maintaining bridge connectivity. Full documentation is served on demand — the AI fetches schemas only when it needs them.
Every existing approach to browser automation has a real cost. Here's the honest comparison:
| Browser Use / Playwright | BrowseAI / Browserbase | Codex / Claude Computer Use | AgentLimb | |
|---|---|---|---|---|
| Setup | Write scripts, manage deps | SaaS config per workflow | Mac only, desktop env required | Copy one prompt |
| Element targeting | CSS/XPath — you maintain them | Visual AI — breaks on updates | Screenshot coordinates — ±1px miss | CDP reads live DOM — semantic |
| Token cost / step | None (pure script) | Cloud fees + AI tokens | 1,000–3,000/screenshot × every step | ~300/step, 85.7% less on hot start |
| Repeat task cost | Fixed | Billed per run | Linear — re-explores every time | Decreasing — muscle memory compounds |
| Login sessions | Extra setup | Cloud — no local sessions | OS-level, unaware of browser state | Your real Chrome — already logged in |
| When site updates | Scripts break, rewrite | Model may degrade silently | Screenshot inference — expensive | AI self-heals muscle selectors |
| Data privacy | Local ✅ | Third-party servers ❌ | Local ✅ | 100% local, 127.0.0.1 ✅ |
| AI choice | Any (pure script) | Platform-dependent | Bundled with Codex / Claude | Any AI that speaks HTTP |
| Shared knowledge | Script = one AI only | Workflow = locked to platform | No persistent memory | Muscle files = cross-AI, transferable |
| Multi-account parallel | Manual orchestration | Platform-dependent | No | ✅ Multiple Chrome profiles, one command |
The unique differentiator: AgentLimb's muscle files live in ~/Desktop/AgentLimb-muscle/ as plain JSON. Knowledge explored by Claude Code today is available to Codex tomorrow — same files, zero re-exploration. Switch AI tools without losing a single learned workflow.
- Marketing — Post to social media, manage campaigns across platforms
- Research — Scrape data, compare products, gather competitive intelligence
- Automation — Fill forms, submit applications, update profiles
- Testing — QA your web app on a real browser with real sessions
- Minimal surface — 16 tools, each does one thing well, composable across any workflow
- Non-invasive — works inside your real browser, not a sandbox
- Local-first — privacy by architecture, not by promise
- AI-agnostic — any tool that can send HTTP can connect; no vendor lock-in
The public repository contains the complete product source:
_locales/ Extension translations
icons/ Extension icons
kernel/ Bridge, CLI, prompt, browser control, and muscle runtime
scripts/ macOS and Windows installers plus release tooling
tests/ Node test suite
ui/ Chrome extension side panel and options UI
website/ Static agentlimb.com site deployed by Cloudflare Pages
忽略上传/ Local-only references, legacy worktrees, private material, and builds
Development commands:
npm test
npm run build # builds into 忽略上传/dist/
npm run release -- 0.2.1 # sets the version and builds release assetsCloudflare Pages should use the repository root as its root directory and website as its build output directory. Release zips are generated in 忽略上传/dist/ and published through GitHub Releases; the entire 忽略上传/ directory is never committed.
- Website: agentlimb.com
- Blog: hooosberg.com/apps/agentlimb
- Windows test checklist: docs/windows-testing.md
- Development notes: docs/development/README.md
- Tutorials: agentlimb.com/tutorials.html
- AI Tools Directory: agentlimb.com/tools.html
- News: agentlimb.com/news.html
- Privacy Policy: agentlimb.com/privacy.html
- Terms of Service: agentlimb.com/terms.html
- License: agentlimb.com/license.html
- GitHub: hooosberg/AgentLimb
- Email: zikedece@proton.me
Built by hooosberg:
- BeRaw — Behance raw-image grabber
- Packpour — App Store Connect locale filler
- WitNote — local-first AI writing companion
- GlotShot — perfect App Store preview images
- TrekReel — outdoor trails, cinematic reels
- DOMPrompter — visualize DOM for AI code
- UIXskills — AI → JSON → Whiteboard → UI
Business Source License 1.1 — free for personal use. Commercial use requires a license. Converts to Apache 2.0 on 2030-04-12.
Copyright © 2025 hooosberg. All rights reserved.