Playwright-backed browser automation for LLMs with a twist: every session starts with a stated goal, and every action (goto, click, fill, press, eval) gets logged with timestamps, post-action URLs, and aria-snapshots. When you call session_end, you get a structured JSON trace on disk that a second LLM can read via qa_get_trace to judge whether the goal was actually met. Five tools, stdio transport, runs on npx with no cloud dependency. Reach for this when you need an audit trail of what your agent did in the browser, not just real-time HTML responses. Each session gets its own Chromium context, and you can toggle headless mode with an env var to watch it work.

A Model Context Protocol server that gives an LLM a real Chromium browser, records every action it takes toward a stated goal, and hands back a structured trace so the LLM (or a second LLM) can decide whether the goal was actually achieved.
Built on Playwright. Five tools, one binary, no cloud dependency.
Most browser-tool MCP servers are stateless — the LLM clicks, gets HTML back, repeats. There's no record of what happened, no way to grade the run after the fact, and no goal context.
cowork-qa-mcp flips that:
goto, click, fill, press, eval) is recorded with timestamps, the URL after, and the page's aria-snapshot.qa_get_trace call.The orchestrating LLM can then reason over the trace ("did this run actually fulfill the goal, or did it click the wrong button?") instead of trusting the run-time chatter.
| Tool | What it does |
|---|---|
session_start | Open a fresh tab, optional starting URL, return a session id |
session_act | Run one of: goto, click, fill, press, eval. Records the step. |
session_observe | Return current URL + full aria-snapshot of the page |
session_end | Close the tab, persist the trace to disk, return the file path |
qa_get_trace | Return the goal, every step, final URL, and final aria-snapshot — formatted for an LLM to read |
Requires Node 20+. The package is on npm — no clone needed.
# Try it once, no install
npx cowork-qa-mcp
# Or install globally
npm install -g cowork-qa-mcp
The first install pulls Chromium via Playwright's postinstall (~150 MB).
claude mcp add cowork-qa --scope user -- npx -y cowork-qa-mcp
To watch the browser instead of running headless:
claude mcp add cowork-qa --scope user \
-e COWORK_QA_HEADED=1 \
-- npx -y cowork-qa-mcp
Verify with /mcp inside a fresh claude session — you should see cowork-qa ✓ connected and 5 tools.
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"cowork-qa": {
"command": "npx",
"args": ["-y", "cowork-qa-mcp"]
}
}
}
Any client that speaks the MCP stdio transport works. Point its server config at npx -y cowork-qa-mcp.
git clone https://github.com/inSideos-designs/cowork-qa-mcp.git
cd cowork-qa-mcp
npm install
npm run build
node dist/server.js # stdio server, expects an MCP client
This server is also published on the official MCP Server Registry as io.github.inSideos-designs/cowork-qa — clients that auto-discover from the registry will find it without any manual config.
| Variable | Default | Purpose |
|---|---|---|
COWORK_QA_HEADED | unset (headless) | Set to 1 to launch Chromium with a visible window |
COWORK_QA_DATA | <cwd>/.cowork-qa | Directory where <session-id>.json traces are written |
A typical end-to-end loop the orchestrating LLM runs:
session_start({ goal: "find the cheapest 14\" MacBook Pro on apple.com",
url: "https://www.apple.com/shop/buy-mac/macbook-pro" })
→ { session_id: "abc-123" }
session_observe({ session_id: "abc-123" })
→ URL + aria-snapshot
session_act({ session_id: "abc-123", action: "click",
target: "button:has-text('Continue')" })
# ... more acts / observes ...
session_end({ session_id: "abc-123" })
→ { steps: 7, trace_path: "~/.cowork-qa/abc-123.json" }
qa_get_trace({ session_id: "abc-123" })
→ Goal: ...
Steps (7 total): ...
Final URL: ...
Final aria-snapshot: ...
Each trace is a JSON file:
{
"session_id": "abc-123",
"goal": "...",
"steps": [
{
"t": 142,
"action": "click",
"args": { "target": "...", "value": null },
"url_after": "...",
"aria_after": "..."
}
],
"final": { "url": "...", "aria": "..." },
"path": "/.../abc-123.json"
}
session_observe calls don't show up in the trace's step count — only session_act calls do. The final aria-snapshot is captured at session_end.eval runs the JS expression but doesn't return the value to the caller — only side effects on the page are observable.button:has-text("Send")), and role= selectors all work.MIT — see LICENSE.
PRs welcome. Keep it small: this is meant to stay a thin, auditable server.
COWORK_QA_HEADEDSet to '1' to launch Chromium with a visible window instead of headless.
COWORK_QA_DATADirectory where session traces (<session-id>.json) are written. Defaults to <cwd>/.cowork-qa.
therealtimex/browser-use
jae-jae/fetcher-mcp
merajmehrabi/puppeteer-mcp-server
com.thenextgennexus/playwright-mcp-server
saik0s/mcp-browser-use