ClaudeMap

browse

Let your AI agents use your real browser — your tabs, your logins, your extensions: the browse MCP server plus a Chromium extension work in background tabs and hand the tab back when it's your turn (a login, 2FA, a payment), continuing once you click Continue (MIT)

MCP Serversbrowser automationmcp serverchromium extensionopencodeplaywright

browse is a browser-automation MCP server plus a Chromium extension from Anomaly (anomalyco, the same org as opencode — anomalyco/opencode is already in this directory; 460 stars as of writing, 2026-11-11): agents get a browse MCP server that opens and drives tabs with Playwright in the background of your daily browser — screenshots, network capture to a redacted HAR, screen recording; when a login, 2FA prompt or payment needs a human, the agent hands the tab back and continues once you click Continue. It works in Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera and Chromium, and the README's support table lists per-agent config paths for 31 agents (OpenCode, Claude Code, Claude Desktop, Cursor, Codex, VS Code, Zed and more). The design stays out of your way: agent tabs are grouped and collapsed, idle ones auto-close after 10 minutes, Patchright hides the usual automation tells and input moves like a person. The security model includes a relay token and per-browser pairing keys (ECDH exchange, HMAC proof on every connection), a restricted execute realm, per-session domain allowlists, and site secrets (passwords / TOTP seeds) that never enter transcripts, logs or HARs. MIT; no pull requests — open an issue.

What it does

  • Agents open and drive background tabs in your daily browser with Playwright (Patchright hides the usual automation tells); tabs group out of your way and idle ones auto-close after 10 minutes
  • Human handoff: the agent hands the tab back for logins, 2FA or payments and continues once you click Continue — only a real click finishes a handoff, so a page can't fake it
  • Network capture to a redacted HAR, tab recording to MP4/WebM, and a flight recorder that keeps the last minute of video for when something breaks
  • Site secrets: stored passwords and TOTP seeds are typed only into matching origins and never enter the execute realm, results, logs or HARs
  • Per-session domain allowlists, a relay token plus per-browser pairing keys (ECDH exchange, HMAC proof on every connection), and a restricted execute realm
  • The support table covers 31 agents with their user-level config files — Claude Code, OpenCode, Claude Desktop, Cursor, Codex, VS Code, Zed and more

When to use it

For automation that needs your real logged-in session: working inside SSO-only internal tools, filling checkout forms up to the payment step (which hands back to you), checking a web UI under your account, recording demos of real page state — and letting a coding agent verify the page it just changed.

Getting started

# installer (puts the browse command in ~/.local/bin and runs browse install; reads the channel manifest from Anomaly's release service and verifies size and SHA-256):
curl -fsSL https://opencode.ai/browse/install | bash

# Windows (PowerShell):
irm https://opencode.ai/browse/install.ps1 | iex

# with Node.js (installs @opencode/browse globally so agents run browse from PATH):
npx @opencode/browse install

# any other stdio MCP client — add the browse command with the mcp argument (verbatim from the README):
{ "mcpServers": { "browse": { "command": "browse", "args": ["mcp"] } } }

# requirements: a Chromium browser, and opencode or another MCP client; the npx install also needs Node.js 22.19+
# during install, pick your agents in the picker, then add the extension from the Chrome Web Store page it opens

Things to know

Updates are automatic by default: binary installs check their channel's manifest at launch and at least once an hour, npm installs check npm's dist-tag once an hour — set OPENCODE_BROWSER_AUTO_UPDATE=off or choose Don't check on the extension's settings page to stop it. Connections across machines are not encrypted by browse — use Tailscale or WireGuard, not the open internet. Stealth is best-effort, not guaranteed: the browser shows its 'started debugging' bar while the debugger is attached, focus emulation keeps background agent tabs visible, and pages can detect the ghost cursor while it is shown. Chrome's own pages (the Web Store included) can't be lent to agents, and file pages need 'Allow access to file URLs' on browse's details page. The restricted execute realm shares the relay's process — treat code an agent runs as code running as you. Bot challenges are detected and reported, never solved. Contributions: no pull requests; open an issue instead. MIT. 460 stars as of writing (2026-11-11).

Repository

anomalyco/browse

Charted

Related on the map

  • Blitz Strike

    MCP Server

    A penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.

    • mcp-server
    • security
    • pentest
    • methodology
  • Ripwire

    MCP Server

    Red Hat's "ripgrep for AI context" — a C++23 zero-dependency CLI that also speaks MCP.

    • mcp
    • code-context
    • cli
    • red-hat
  • Deja Vu

    MCP Server

    One shared session memory for 25+ coding agents — Claude Code, Codex, and Cursor finally remember the same project.

    • mcp
    • memory
    • cross-agent
    • context