Skip to main content
Glama

byob

Bring Your Own Browser — let your AI assistant use the Chrome you already have open.

License: MIT MCP Chrome MV3 v0.5

English · 中文


byob is a local MCP server that lets AI coding tools (Claude Code, Cursor, Cline, Windsurf, etc.) directly control your real Chrome — the one where you're already logged into everything.

"read my Twitter timeline and summarize the top 5 posts"
"google 'mcp protocol spec', click the first result, read the page"
"take a screenshot of example.com"
"grab my GitHub session cookie so I can curl with it"
"open my Gmail tab and tell me how many unread"
"fill in this form with my details and submit it"

WebFetch

Headless Puppeteer

byob

Sees logged-in pages

❌

⚠️ manual cookie copy

✅ already logged in

Passes bot detection

❌

❌

✅ real human browser

Setup time

0

hours

~5 min

Cloud cost

free

$$

free

Compared with tools that attach over Chrome's remote-debugging port (DevTools MCP, agent-browser, Playwright --cdp): byob is an extension, so there's no debugging port to open, no "Allow remote debugging?" prompt on every connection, and nothing else on your machine can hijack your logged-in browser through it.


Install

curl -fsSL https://raw.githubusercontent.com/wxtsky/byob/main/install.sh | bash

On Windows, run the same command from Git Bash or MSYS2. The script checks prerequisites (Node.js ≥ 20, bun, Chrome/Edge/Brave), clones the repo, builds everything, and walks you through MCP registration interactively. If bun is not installed, it offers to install it for you using the native installer for your OS.

Set BYOB_INSTALL_DIR to change the install location (default: ~/byob). Advanced: BYOB_REPO, BYOB_REF, and BYOB_SKIP_SETUP=1 are supported for forks, pinned refs, and CI-style dependency install only.

Manual install

Requires Node.js ≥ 20, bun, Chrome, and any MCP-compatible AI tool.

git clone https://github.com/wxtsky/byob
cd byob
bun install
bun run setup

bun run setup walks through the install interactively:

  1. Pick output language (English / 中文)

  2. Generates a unique extension key for you

  3. Builds the Chrome extension

  4. Writes the config that lets Chrome talk to byob

  5. Prompts you to multi-select your AI tools. Claude Code gets the bundled plugin (Skill + auto-started MCP); other clients get their MCP config.

After the script finishes, three manual steps remain:

Step 2 — Load extension in Chrome

Open chrome://extensions in Chrome.

  1. Top-right → turn ON Developer mode

  2. Top-left → click Load unpacked

  3. Select the folder printed in your terminal, something like:

    /your/path/to/byob/packages/extension/output/chrome-mv3

Step 3 — Restart Chrome

Quit Chrome completely (⌘Q on Mac / close all windows on Windows), then reopen.

Closing a single tab or window is not sufficient — Chrome only reads the Native Messaging config at startup.

Step 4 — Claude Code plugin / manual MCP reference

The setup script registers your selected tools automatically. The block below is for reference only — use it if you skipped the prompt or want to register a different tool later:

claude plugin marketplace add wxtsky/byob
claude plugin install byob@byob --scope user

Run /reload-plugins in Claude Code after installing. The plugin includes the /byob:control-chrome Skill and starts its bundled MCP server automatically; do not also register a second byob MCP server.

For local development without installing:

claude --plugin-dir /path/to/byob/plugins/byob
claude mcp add byob -s user -- /path/to/tsx /path/to/byob-mcp.ts

To enable browser_evaluate, add -e BYOB_ALLOW_EVAL=1 after -s user.

codex mcp add byob -- /path/to/tsx /path/to/byob-mcp.ts

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "byob": {
      "command": "/path/to/tsx",
      "args": ["/path/to/byob-mcp.ts"]
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json (same JSON format as Cursor):

{
  "mcpServers": {
    "byob": {
      "command": "/path/to/tsx",
      "args": ["/path/to/byob-mcp.ts"]
    }
  }
}

Open Cline sidebar → MCP Servers → Configure, then add (same JSON format):

{
  "mcpServers": {
    "byob": {
      "command": "/path/to/tsx",
      "args": ["/path/to/byob-mcp.ts"]
    }
  }
}

The actual paths are printed by the setup script. The examples above use shortened paths for readability.
To enable browser_evaluate, add "env": { "BYOB_ALLOW_EVAL": "1" } to the config (or -e BYOB_ALLOW_EVAL=1 for CLI tools).

Step 5 — Wait for setup to confirm the bridge is online

After you finish steps 2 and 3 (load the extension and ⌘Q-restart Chrome), setup auto-detects the bridge coming online and prints ✓ bridge online. The installation is complete — open a new session in your AI tool and try "use byob to read ...".

If you exited setup early with Ctrl+C, or want to check the state later:

bun run doctor

bun run doctor prints actionable fixes under every ✗ (e.g. "⌘Q-restart Chrome", "extension ID mismatch", etc.).


Related MCP server: agent-browser-mcp

Tools

v0.5 has 25 tools. Elements are targeted by refs from the snapshot ([ref=e12]), a CSS selector, or x/y.

Tool

What it does

browser_snapshot

The page as an accessibility outline with refs — interactive-only, diff, scoped; iframes (even cross-origin) inline

browser_find

Find by role+name / text / label / placeholder / testId, optionally click/fill/check right away

browser_click

Click (real mouse events), double/right click, modifiers; auto-waits and reports navigation / new tabs / dialogs

browser_type

Fill inputs, textareas, contenteditable (React/Vue-safe); per-key mode; submit

browser_press_key

Keys and combos with real key codes (Enter, Shift+Tab, ControlOrMeta+A)

browser_hover / browser_select / browser_scroll / browser_drag

Menus, <select>, page or container scrolling, HTML5 drag-and-drop

browser_upload_file

Attach local files to <input type=file>

browser_batch

Many steps in one call (fill a whole form and submit)

browser_wait

Wait for element state, text, URL, load / network idle

browser_read

Content as text / markdown (reader mode) / html / links / tables, with outline and filter

browser_screenshot

Viewport / full page / element / clip → image returned to the model; ref labels; PDF

browser_navigate

Open a URL in a background tab, back / forward / reload

browser_tabs

List / open / select / close tabs, search history

browser_console

Console messages and page errors, buffered

browser_network

Record traffic (summary / JSON / HAR) or intercept: block / mock / modify

browser_inspect

Element box / state / styles, Web Vitals

browser_storage

Cookies and local/session storage — get / set / clear

browser_dialog

Answer alert / confirm / prompt explicitly

browser_emulate

Device, dark mode, geolocation, timezone, locale, offline, headers

browser_clipboard

Read / write clipboard text

browser_download_images

Save a page's images

browser_evaluate

Run page JavaScript (only with BYOB_ALLOW_EVAL=1)

Schemas: shared/src/tools.ts. Upgrading from 0.4? Tool names changed — see the changelog.


How it works

AI tool → byob-mcp → byob-bridge → Chrome extension → your tab
         (stdio)    (Unix socket) (Native Messaging) (Chrome DevTools Protocol)
  • byob's own page scripts run in an isolated world: they share the DOM but not JavaScript with the page — no window.__byob*, no data-byob-* attributes, nothing a site can see or tamper with.

  • Refs live in the extension's memory, bound to Chrome's backend node ids; they survive re-renders and are refused (never silently re-targeted) after navigation.

  • Each MCP client session has its own working tab; tabs the agent opens are grouped under byob and never steal focus.

All communication stays local. No data leaves your machine. When Chrome closes, all byob processes exit automatically.


Everyday commands

bun run setup      # install or re-install
bun run doctor     # check what's working
bun run bridges    # list running bridge processes
bun run logs       # tail the bridge log
bun run unsetup    # remove everything
bun run policy     # show / change which sites byob may touch (flags below)

All run from the byob repo root.


Reliability

  • Auto-waiting actions. Before clicking or typing, byob waits until the element is attached, visible, stable, enabled and actually on top at its center; errors say what's wrong ("covered by <div#cookie-banner>").

  • Actions report their effects — navigation (waited for), same-page URL changes, new tabs, JS dialogs — so the agent doesn't have to poll.

  • Dialogs never hang a call. A click that opens confirm() returns immediately with the dialog text.

  • End-to-end cancellation. Ctrl+C propagates through the entire chain (MCP → bridge → extension → Chrome).

  • Sleep/wake recovery. After a laptop sleep cycle, byob resets its debug sessions so the next call starts clean.

  • Tested end to end on Chrome for Testing: BYOB_E2E=1 bun test e2e (see CONTRIBUTING).


Security

  • browser_evaluate (and browser_wait's fn) are off by default — enable with BYOB_ALLOW_EVAL=1. Every call is audit-logged.

  • chrome://, file://, Google/MS/Apple login pages are blocked by default.

  • Per-site allow/deny lists, set by you — never by the agent:

    bun run policy --deny "**.chase.com" "mail.proton.me"   # never touch these
    bun run policy --allow "**.github.com"                  # or: only these
    bun run policy --clear-allow --clear-deny                # reset
    bun run policy --allow-file on                           # permit file:// URLs

    Patterns: example.com (exact), *.example.com (subdomains only), **.example.com (apex + subdomains), * (everything). Deny wins over allow. A non-empty allow list means allow-list-only. Enforced when byob attaches to a tab, so it covers every tool.

  • JS dialogs are never auto-accepted; a confirm stays open until the agent answers it explicitly.

  • Credential fields are redacted. Values in password / OTP / card / email inputs are never sent to the model; they surface as [redacted].

  • Each install gets a unique extension key — no collisions.

  • Socket files are 0600, dirs are 0700. Other users can't see them.

  • Zero outbound network traffic. No analytics, no pings, no crash reports.

  • Chrome displays a "byob is debugging this browser" banner on active tabs. This is a Chrome security feature and cannot be suppressed.


Troubleshooting

Symptom

Cause

Fix

No live bridge

Chrome not running or extension disabled

Check chrome://extensions

cdp_attach_failed

DevTools open on that tab

Close DevTools

url_forbidden

Blocked by your site policy

bun run policy

stale_ref

The page navigated since the snapshot

Snapshot again

element_covered

A banner / modal is on top

Close it, or force:true

extension_not_connected

Extension lost connection

Reload at chrome://extensions

Nothing works after install

Chrome was not fully restarted

Quit Chrome completely (⌘Q) and reopen

Run bun run doctor for detailed diagnostics on which step failed.


Platform notes

Platform

Auto

Manual

macOS

Auto-registers selected MCP tools

Open chrome://extensions and load the unpacked extension

Windows

Same + writes Native Messaging host to registry

Same as macOS

Linux

Auto-registers selected MCP tools

Same as macOS


More

MIT licensed. byob has broad access to your browser — only use it on machines and accounts you own.

Maintenance

ActivityMaintained
ResponsivenessUnresponsive

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI assistants to control and automate your Chrome browser directly, leveraging existing login states and configurations for tasks like content analysis, semantic search across tabs, screenshots, network monitoring, and interactive operations.
    12
    MIT
  • A
    license
    B
    quality
    F
    maintenance
    Enables AI agents to directly control your real Chrome browser with full context including login sessions, cookies, and open tabs. It provides tools for page scanning, JavaScript execution, CDP control, screenshots, and physical mouse/keyboard input for authentic browser automation.
    20
    245
    MIT
  • A
    license
    Not graded
    quality
    D
    maintenance
    Enables AI assistants to control a Chrome browser through natural language, supporting navigation, interaction, tab management, and data extraction.
    MIT
  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI assistants and terminals to control a user's already logged-in Chrome session, allowing them to navigate pages, read content, click and type, run JavaScript, and inspect console or network activity.
    30
    MIT