Remote Browser + MCP Tool Gateway (Multi-Site MVP)

This project provides a minimal HITL (human-in-the-loop) remote browser session and an MCP tool gateway for restricted search/extraction. It currently supports XHS and includes real search adapters for Yelp and TripAdvisor.

Requirements

• Node.js >= 18

• Playwright Chromium (npx playwright install chromium)

Install

npm install
npx playwright install chromium

Copy the environment template and adjust as needed:

cp .env.example .env

Run

npm run dev

This starts:

• HTTP server on http://HOST:PORT

• MCP server on stdio (connect with an MCP client)

Docker (single-user + noVNC)

This path is intended for a single user (or MAX_SESSIONS=1). It runs the browser inside Xvfb and streams the desktop via noVNC.

Build:

docker build -t browser-mcp-demo .

Run:

docker run --rm \
  -p 3000:3000 -p 7900:7900 \
  -e HOST=0.0.0.0 \
  -e HEADLESS=false \
  -e MAX_SESSIONS=1 \
  -e VIEW_MODE=novnc \
  -e PUBLIC_BASE_URL=http://YOUR_SERVER_IP:3000 \
  -e NOVNC_URL_TEMPLATE="http://YOUR_SERVER_IP:7900/vnc.html?autoconnect=1&resize=scale&path=websockify" \
  -e PROFILES_DIR=/data/profiles \
  -e AUDIT_LOG_PATH=/data/logs/audit.log \
  -e DELETE_PROFILE=false \
  -v "$PWD/profiles:/data/profiles" \
  -v "$PWD/logs:/data/logs" \
  browser-mcp-demo

Notes:

• VIEW_MODE=novnc makes /session/view/:id embed the live browser stream.

• Update PUBLIC_BASE_URL and NOVNC_URL_TEMPLATE with your public host or domain.

HITL Login Flow

1. Call MCP tool create_session -> { sessionId, viewUrl }

2. Open viewUrl in your browser.

   • Default mode: a local Chromium window is opened for login.

   • noVNC mode (VIEW_MODE=novnc): the remote browser stream is embedded in the page.

3. Login on that window (QR/OTP/2FA handled by the user).

4. Call wait_for_login until status is READY (site-aware when site is provided).

MCP Tools (stdio)

Tools:

• create_session

• wait_for_login (optional site)

• platform_search (site-aware via site param)

• xhs_open_and_extract (site-aware via site param)

• destroy_session

Example (pseudo):

const session = await client.callTool("create_session", {});
await client.callTool("wait_for_login", { sessionId: session.sessionId, timeoutSec: 120 });
const results = await client.callTool("platform_search", {
  sessionId: session.sessionId,
  query: "camping",
  maxNotes: 10,
  scrollTimes: 0,
  site: "xhs" // xhs | yelp | tripadvisor
});
const detail = await client.callTool("xhs_open_and_extract", {
  sessionId: session.sessionId,
  url: results.notes[0]?.url,
  site: "xhs"
});

Security Boundary

• Tools return sanitized structured JSON only.

• No cookies, localStorage, sessionStorage, storageState, or userDataDir exposure.

• No screenshot tool.

• Audit log is written to logs/audit.log with redaction.

Agent HTTP Endpoints

• POST /agent/run → start a run and execute until login required or done

• POST /agent/continue → continue a run after user login

• GET /agent/run/:id → fetch current run state

Example request body:

{
  "query": "camping",
  "maxNotes": 10,
  "scrollTimes": 0,
  "detailCount": 3,
  "detailParallel": 4,
  "site": "xhs"
}

Configuration

Key environment variables:

• HOST, PORT, PUBLIC_BASE_URL

• UI_DIST_DIR (serve built UI from the same server)

• VIEW_MODE (info | novnc)

• NOVNC_URL_TEMPLATE (supports {sessionId} placeholder)

• OPENAI_API_KEY, OPENAI_MODEL

• AGENT_RUN_TTL_MINUTES

• MAX_SESSIONS, SESSION_TTL_MINUTES

• PROFILES_DIR, DELETE_PROFILE

• HEADLESS

• XHS_BASE_URL

• AUDIT_LOG_PATH

Notes

• XHS, Yelp, and TripAdvisor all support search in the current adapter layer. XHS detail extraction is implemented; Yelp/TripAdvisor detail extraction is still stubbed.

• The DOM selectors for each site may change. Update src/browser/xhs.ts or src/sites/*.ts if extraction breaks.

• This MVP does not implement large-scale crawling or anti-bot bypass.