Skip to main content
Glama
chrome-devtools-mcp-broker

chrome-devtools-mcp-broker

by ccomkhj

chrome-devtools-mcp-broker

Let several AI agents use Chrome at the same time, each in its own browser.

Why

Google's chrome-devtools-mcp can serve several agents from one server, but they all share one browser:

  • Agents see each other's pages. isolatedContext separates cookies and storage, but every agent's list_pages shows the other agents' pages with titles and URLs, and evaluate_script can run on any of them.

  • Agents wait for each other. Browser-wide tools such as evaluate_script, new_page or a performance trace run one at a time; while one runs, no other agent's call can start.

  • Large results flood the context. A big snapshot or listing comes back in full.

This broker gives each agent session its own chrome-devtools-mcp worker and temporary Chrome profile, so a session sees only its own pages and never waits for another session's call. It caps each reply at a character budget and caches the full result, so agents can page or filter it without running the action again. It drives the published chrome-devtools-mcp through MCP, unmodified. Evidence: docs/VALIDATION.md.

Related MCP server: SpidexLab's Multi-Tenant Browser MCP

When to use it

You have

Use

Several agents on one MCP server or client, e.g. parallel subagents

This broker

Large snapshots or listings you want to page or filter

This broker

Agents in separate processes, each starting its own server

Upstream with --isolated

Agents that may share pages and mostly navigate, click and fill, or tight memory

Upstream with isolatedContext

Your own logged-in Chrome (--autoConnect, --browser-url)

Upstream

Quick start

Requires Node.js 22.12+ and Google Chrome (or Chromium via --chrome-path).

(umask 077; openssl rand -hex 32 > ~/.chrome-broker-token)   # once
CHROME_BROKER_TOKEN=$(cat ~/.chrome-broker-token) npx -y chrome-devtools-mcp-broker
# listening at http://127.0.0.1:7331/mcp

claude mcp add --transport http chrome-broker http://127.0.0.1:7331/mcp \
  --header "Authorization: Bearer $(cat ~/.chrome-broker-token)"

Other clients: see examples/mcp-config.json. Or let the client start its own broker over stdio (no token; one session limit per client):

{ "mcpServers": { "chrome-broker": { "command": "npx", "args": ["-y", "chrome-devtools-mcp-broker", "--transport", "stdio"] } } }

Tools

Tool

Purpose

chrome_session

open a session (one per agent), info, or close it

chrome_tools

List, search, or get the input schema of upstream tools; does not launch Chrome

chrome_call

Run an upstream tool (tool, args) in your session; returns bounded text and a resultId

chrome_read

Page (offset) or filter a cached result by resultId

Output options for chrome_call and chrome_read: maxChars (reply budget, default 6,000 characters), filter (case-insensitive line match), select (JSON Pointer into the raw result), includeImages.

chrome_session {"action":"open"}                                      -> {"sessionId":"S",...}
chrome_call    {"sessionId":"S","tool":"new_page","args":{"url":"https://example.com"}}
chrome_call    {"sessionId":"S","tool":"take_snapshot","args":{}}     -> first part, "resultId":"R", "nextOffset":N
chrome_read    {"sessionId":"S","resultId":"R","filter":"heading"}    -> only heading lines
chrome_read    {"sessionId":"S","resultId":"R","offset":N}            -> next page, no re-run
chrome_session {"action":"close","sessionId":"S"}

Page tools act on the session's selected page, so they take no pageId. Calls in one session run in order; different sessions run in parallel.

Configuration

npx -y chrome-devtools-mcp-broker --help lists every option and default. Flags after -- go to every upstream worker:

npx -y chrome-devtools-mcp-broker --max-sessions 8 -- --viewport=1280x720

Flags that would break per-session isolation (attaching to an existing browser, a shared profile, and so on) are rejected at startup; see deniedFlags in src/upstream.js.

Security

For one user's trusted agents on their own machine. HTTP binds to 127.0.0.1 and requires the token. It is not a sandbox: pages run JavaScript and Chrome runs as your OS user. Threat model and reporting: SECURITY.md.

Limitations

  • Each session runs its own worker and Chrome, so memory grows with sessions and tabs. Size --max-sessions to fit.

  • A call that times out closes its session; the broker never retries.

  • Sessions and cached results live in memory; cached results expire after 10 minutes.

  • Agents run in parallel only if the MCP client sends their calls concurrently.

  • Verified on macOS only so far.

More

  • docs/ARCHITECTURE.md: design and trade-offs

  • docs/VALIDATION.md: measurements and how to reproduce them

  • Development: npm ci && npm test (no Chrome needed); CHROME_PATH=/path/to/chrome npm run test:chrome for real-Chrome tests

License

MIT. Independent project, not affiliated with Google.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Lets several AI clients drive one shared, signed-in Chrome at the same time, giving each session its own background tabs and tab group with per-tab leases so mutations queue while reads overlap. Exposes a full Chrome DevTools-style toolset plus Claude's 31-member browser toolset, along with an HTTP endpoint that turns Messages API tool_use blocks into tool_result blocks.
      MIT