agentic-browser

mcp
Security Audit
Fail
Health Pass
  • License — License: GPL-3.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 14 GitHub stars
Code Fail
  • network request — Outbound network request in agents/react_tools.py
  • new Function() — Dynamic code execution via Function constructor in clients/browser-extension/entrypoints/background.ts
  • network request — Outbound network request in clients/browser-extension/entrypoints/sidepanel/App.tsx
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

AI agent for web automation and scraping, designed to plan, execute, and verify actions across any website

README.md

Agentic Browser - Model-Agnostic Web Automation

The Open Agent Browser Extension Powered by Python & MCP

Mission: Build an intelligent browser agent that doesn’t just understand the web — it acts on it. Fully model-agnostic, privacy-respecting, and BYOKeys-ready.


Overview

Agentic Browser is a next-generation browser extension powered by a Python MCP (Model Context Protocol) server that bridges modern LLM reasoning with real browser interactivity.

Unlike typical AI assistants, this agent:

  • Understands complex web content via real-time DOM inspection.
  • Takes actions such as filling forms, navigating, comparing data, and executing custom scripts.
  • Adapts to any model backend — OpenAI, Anthropic, Ollama, local LLaMA, Mistral, and more.

Architecture

Agentic Browser Architecture

Key Principles

  • Model-Agnostic: Seamlessly switch between LLM providers (OpenAI, Anthropic, Ollama, LM Studio).
  • BYOKeys: Full privacy. Your keys, your local context, no vendor lock-in.
  • MCP-Compliant: Leverages the Model Context Protocol for secure, structured tool interaction.
  • Safe Execution: Declarative action system with mandatory user approval for browser operations.

Technical Stack

Component Technology Role
Frontend React, Vite, WXT High-performance, intelligent sidepanel UI
Agent Core LangGraph, LangChain Orchestration of complex multi-step workflows
Backend Python (FastAPI), MCP Core logic, tool execution, and state management
Browser Control WebExtensions API Secure DOM manipulation and tab management
Safety Permission Guards Per-action approval and comprehensive logging

Monorepo Layout

clients/
  browser-extension/   # WXT browser extension client
  debug-web/           # Vite debug dashboard client
  telegram-bot/        # Future Python Telegram client

agents/
core/
mcp_server/
models/
routers/
services/
main.py                # Root Python backend entrypoint
pyproject.toml         # Root Python package metadata
pnpm-workspace.yaml    # Frontend workspace definition

Workspace Commands

pnpm install
pnpm dev:debug
pnpm dev:extension
pnpm build:debug
pnpm build:extension

The Python backend stays at the repository root and is still managed through pyproject.toml and uv.lock.


Use your browser from Claude Code, Codex, T3 Code and other MCP clients

The extension can be driven by any coding agent that speaks MCP. The agent works in your real Chrome, with your logins, through a small local bridge.

Claude Code / Codex / T3 Code / Cursor / OpenCode
        │  MCP (stdio, or HTTP at 127.0.0.1:5455/mcp)
        ▼
agentic-mcp  ──  bridge on 127.0.0.1:5455
        ▲  WebSocket /bridge/ws (the extension dials out)
        │
Agentic Browser extension  ──  chrome.debugger + content scripts  ──  your tabs

Setup takes three steps:

  1. Build the extension (pnpm build:extension) and load clients/browser-extension/.output/chrome-mv3 as an unpacked extension in Chrome.
  2. Install the MCP server. You don't need the full backend for this:
    uv tool install ./mcp_server     # only the MCP server and its few dependencies
    # or, inside a full checkout: uv sync
    
  3. Register it with your agents:
    agentic-mcp install              # finds Claude Code, Codex, Cursor, Gemini CLI, OpenCode
    agentic-mcp install --dry-run    # show the changes first
    agentic-mcp config               # print snippets to paste by hand
    
    T3 Code runs Claude Code and Codex, so it uses whatever you register for those two.

Each MCP client starts agentic-mcp over stdio. The first one to start binds 127.0.0.1:5455 and becomes the bridge. Any later ones relay through it, using a token in ~/.agentic-browser/bridge-token. The backend (agentic-api-run) also starts the bridge when it boots, so the extension and backend can run together. To keep an HTTP endpoint up for clients that only support URLs, run agentic-mcp serve.

The extension's settings panel has an "MCP Bridge" section with the connection status and a URL override. agentic-mcp status shows the same thing from the terminal.

Tools

There are four tools, on purpose:

Tool What it does
browser_tabs List, open, close or focus tabs. Tabs the agent opens go in an "Agent" tab group.
browser_snapshot Accessibility outline of a tab with [ref=e12] handles, including iframes (f3e1). Can add a screenshot.
browser_act One action on a ref, CSS selector or text=Label: click, type, press, select, check, hover, scroll, navigate, back, forward, reload, wait_for, evaluate, or answering a dialog.
browser_execute Code mode. Runs a JavaScript program against a browser / tab API, so ten steps cost one tool call and only the data you return goes back to the model.

browser_execute runs in a QuickJS sandbox inside agentic-mcp. Scripts get no network, file system or Node APIs. Each burst of JS between browser calls is capped at 5 seconds, which kills infinite loops. A state object carries data between calls.

const t = await browser.open("https://news.ycombinator.com");
return await t.evaluate(() =>
  [...document.querySelectorAll(".titleline > a")].slice(0, 5).map(a => a.textContent));

Set AGENTIC_BROWSER_TOOLS=execute to expose only browser_tabs and browser_execute, or plain to hide browser_execute.

Clicks, typing and key presses go through chrome.debugger, so pages see trusted input events. While it's attached, Chrome shows a "started debugging this browser" bar, which goes away after 90 seconds of inactivity. If the debugger can't attach (DevTools is open, or you're in Firefox), the extension falls back to DOM events.

Security

  • The bridge listens only on 127.0.0.1.
  • WebSocket connections are accepted only from chrome-extension:// and moz-extension:// origins. To allow only your build, set AGENTIC_BROWSER_EXTENSION_IDS to a comma-separated list of extension IDs.
  • /mcp rejects cross-site Origin and foreign Host headers, which blocks DNS rebinding.
  • The relay endpoint needs the token in ~/.agentic-browser/bridge-token (file mode 0600).

Slash Commands & Tools

Command Capability
/browser-action Execute UI automation (click, type, scroll, navigate)
/voice Enter voice interaction mode
/react-ask Engage with the ReAct reasoning agent
/google-search Rapid web search via multi-provider adapters
/gmail-unread Check and summarize your unread emails
/calendar-events View and manage your upcoming schedule
/youtube-ask Interactive Q&A with video transcripts
/pyjiit-attendance Check academic portal attendance

Roadmap

  • Offline Mode: Local embedding-based RAG using lightweight models - Gemma 4 uncensored
  • Light Mode: For Presemtation
  • Skills: Open Skill Support for stuff like PPT and XLSX
  • Grpaph RAG: For important documents like Resume and Gmail mails
  • Memory System: To have persistant Memory and other relevant Facts
  • Github Integration: Right its leading to Context rott so need shit with proper tool call and all will use Pi (hopefully)
  • While True: Tool called Orgestrated model for the agent.
  • Telegram / External Chat Support: use to ask question plus options to switch context or something or atleasr a standalone app for this shit
  • Computer Use: Use computer also instead of just the browser
  • Website / Native App: To use it beyond tradition browser usecases
  • Form Filling / Browser Action: To take actions till the time shit is not resolved - Infinite Loop
  • Multi Modal: Image support
  • Remote Actions: from phone
  • Playwrite: To take actions from the app itself
  • Compaction: adding compaction for long chats
  • Model Blending: Reasoning pipelines that utilize multiple models for verification.
  • Visual Debugger: Real-time DOM element highlighting during agent actions.
  • Workflow Designer: Create and share custom agentic browsing workflows.
  • MCP Support: agentic-mcp drives the extension from Claude Code, Codex, T3 Code and other MCP clients (see above)

Reviews (0)

No results found