AI browser-control extension

AgentLimb

AgentLimb is a Chrome extension plus a local MCP host for universal browser control. Any MCP-capable agent — Claude Code, Codex, Cursor, Tencent WorkBuddy — pairs by copying one prompt with automatic authorization by default, drives multiple Chrome profiles with targeted routing, supports native HTTP Tool RPC calls, and reuses local Markdown site memory (~/.agentlimb/memory) — 100% local.

Platform
Chrome / Edge / Brave / Vivaldi and other Chromium browsers / Standard MCP / Windows + macOS + Linux / Local loopback
Status
Chrome Web Store / Open source / Offline direct
Audience
Developers and heavy AI users who need agents to work with real websites, logged-in sessions, and multi-account workflows.

Product Overview

AgentLimb is positioned as a universal web-control extension for every platform and every AI tool. It is not tied to one vendor: any agent that speaks standard MCP — Claude Code, Codex, Cursor, Trae, Windsurf, Tencent WorkBuddy, or any model that can run local commands — drives your real browser through it, locating elements, filling forms, clicking, taking screenshots and extracting data on the sessions you are already logged into. No headless browser, no re-login.

0.6.8 introduces a smooth multi-step task lifecycle with anti-flicker tab marking: decoupling single-tool execution from the overall task duration keeps tabs stably inside the orange 'AgentLimb' group with '⏳ AgentLimb' across thinking gaps and sequential steps, completely eliminating flickering. Combined with dynamic multi-terminal matching (devices disappear instantly on disconnect) and smooth auto-ungrouping upon task completion (say/stop), background focus isolation ensures users stay uninterrupted.

Targeted multi-profile routing, native HTTP Tool RPC (POST /api/tool), site memory playbooks, and auto-authorization are fully preserved with dynamic port probing (8765–8770) and Python client integration.

Under the hood it is a zero-dependency Node host plus one MV3 extension: local SSE push instead of polling, per-tab persistent CDP attach, an accessibility tree resident from document_start, adaptive JPEG screenshots, O(1) single-file site memory lookups (~10ms), and code-level credential redaction, guarded by 113 unit and protocol tests passing 100%. 0.7.0 upgrades same-client fast takeover and lock inheritance, resets sticky tabs on client switch, and expands the bridge port pool; Windows users please unzip to a plain-English path.

Pairing & auth
< 1s · Auto-authorized Connecting hosts are trusted automatically by default (strict copy-window mode optional in Advanced); a second AI is discovered in under a second.
Multi-profile
Profile-aware routing Optional browser argument targets specific profiles by label, email substring, or ID with automatic disambiguation.
HTTP Tool RPC
POST /api/tool · Fast Zero-stdio local HTTP RPC for external Python/Node automation scripts, resilient against port drift.
Engineering guards
113 automated tests Full coverage across wire protocol, multi-browser routing, smooth multi-step lifecycle, HTTP RPC endpoints, and site memory chaos suites.

Why It Matters

Each product page is written for users, reviewers, and builders: what it solves, what it does not solve, and what can be reused.

Pluggable site memory that compounds

Exploring a site's DOM for the first time is expensive. Site memory saves direct URLs, selectors, confirmation pitfalls, and verification checks into a local Markdown folder, and automatically nudges the AI in navigate / tab_new and say results. Decoupled from the extension, no Skill to install, reusable by every AI tool.

Install to authorize, one prompt to set up

The prompt carries the extension ID fingerprint so the AI locates the source folder, installs ~/.agentlimb, and wires up MCP on its own. Connecting AIs are trusted automatically by default — no confirmation page, no pairing code. Switch to the 30-minute copy window anytime in Advanced settings.

Reply-window side panel & minimal busy indicator

The side panel is where you talk to the AI: replies arrive via say, and the header simply shows a dot plus 'Working'. In the browser, only the orange tab-group chip shows '⏳ AgentLimb' alongside the in-page glow and ghost cursor — no flashing toolbar badges, no mutated tab titles.

Several AIs, one controller at a time

Up to six MCP hosts connect at once, with trust, chat and drafts isolated per AI. Only one agent runs browser work at any moment: a busy switch is refused rather than forced, and queued calls that expire are dropped — a late browser action is more dangerous than a failed one.

Core Experience

Site memory (site_memory)

Host-side implementation (host/memory.mjs) stored at ~/.agentlimb/memory by default, configurable or toggleable in side panel settings. One PLAYBOOK.md per domain (≤120 lines), auto-injected in tool results; built-in domain allowlist, password/Token/Bearer redaction, deduplication, directory lock, and stale marking after 2 consecutive failures.

One-prompt setup & auto-authorization

Copy the prompt to your AI: it resolves the unpacked extension folder from the extension ID fingerprint, installs the host, and configures MCP. Auto-authorize is on by default; any trusted AI can be removed at any time under Connected AIs.

Any MCP agent

Standard MCP over stdio only: Claude Code, Codex, Cursor, Trae, Windsurf, Tencent WorkBuddy, or any model that can run local commands. The installer only deploys ~/.agentlimb and prints the mcpServers block; enabling and trusting MCP stays with each app.

The side panel is the AI's reply window

Once paired, the panel shows the current AI's replies and connection state. Every reply arrives through say, and panel messages automatically attach a text snapshot and compressed screenshot of the active tab.

Real browser, background tabs & minimal hourglass

The agent operates the Chrome you are using, with logins, cookies and multiple profiles intact. New tabs open in the background inside the orange 'AgentLimb' tab group — which turns into '⏳ AgentLimb' while busy — driven through simulated focus.

17 standard tools & engineering guards

tabs_context, read_page, find, computer, form_input, navigate, javascript_eval, browser_batch (up to 20 actions), read_console_messages, read_network_requests, wait, say, site_memory, ping and more — backed by 113 unit tests and pairing/panel/memory chaos suites.

Design Principles

Store knowledge, drop the transcript

Action-by-action logs are never read twice and slow down search. Site memory keeps only an 8-section playbook with O(1) single-file reads (~10ms), while one-line session logs stay out of the hot path.

Turn prompt rules into code invariants

Soft prompts can be forgotten or skipped by lazy tool loading. Playbooks ride inside navigate/say tool results, while path safety, secret redaction, size caps, and concurrency locks are enforced in host code.

As simple as possible

Install-to-authorize removes extra clicks; busy state only updates the '⏳ AgentLimb' tab-group chip without noisy banners or icon badges; the side panel stays a pure conversation window.

Local-first & serial control

The host listens on 127.0.0.1 only; trust, chat records, site memory and page data never leave the machine. Multiple AIs share one serial controller without mid-task preemption.

Build Diary

The AgentLimb diary covers the whole arc: the 0.2 era of runtime bootstrapping, the 0.5 rewrite around copy-to-pair and a reply-window side panel, and the 0.6 evolution with pluggable site memory, auto-authorization, and minimal tab-group indicators.

Browser automationMCP toolchainsSite memory playbooksAI operating real websites
View all build notes

Privacy and Support

The extension and the local host run entirely on your machine. The MCP host listens only on 127.0.0.1, pairing trust and chat records live in the extension's private storage, site memory stores plain Markdown playbooks with automatic credential redaction, and page data is never uploaded — no cloud, no tracking.