Bright Data MCP browser tools hang 5 min in Docker/Render but work locally

Question

I'm using brightdata-mcp with pydantic-ai and FastAPI inside Docker (and also on Render). Any scraping_browser_* tool makes run_mcp_servers() block for exactly 5 minutes after the model returns, while the exact same code runs instantly on macOS. Non-browser tools like search_engine work fine. What causes the hang and how can I fix it in a containerized environment?

Originally asked on GitHub: https://github.com/brightdata/brightdata-mcp/issues/11

AgentArmor team · Accepted Answer

Status: closed/explained. Bright Data maintainers confirmed the hang is not in the Bright Data MCP server itself but in how pydantic-ai's `run_mcp_servers()` shuts down the MCP child process. In Docker/Render the child receives SIGTERM on shutdown and keeps the WebSocket open until a 300-second watchdog closes the browser tab; locally on macOS the process is force-killed with SIGKILL, so you never notice the timeout.

The fix that worked: launch the MCP server once at app startup and keep it alive across requests instead of calling `glassdoor_agent.run_mcp_servers()` inside every request. Use FastAPI's lifespan context manager:

```python
@asynccontextmanager
async def lifespan(app: FastAPI):
    servers_ctx = glassdoor_agent.run_mcp_servers()
    await servers_ctx.__aenter__()
    try:
        yield
    finally:
        await servers_ctx.__aexit__(None, None, None)

app = FastAPI(lifespan=lifespan)
```

Then in each endpoint just call `await glassdoor_agent.run(prompt)` without wrapping it in `async with run_mcp_servers()`. Also make sure Playwright is installed in your Docker image (`RUN npx -y playwright@^1.51.1 install --with-deps`); without it the container can hang indefinitely instead of waiting 5 minutes. As general advice, avoid starting/stopping MCP servers per request – reuse a single server instance for the app's lifetime.

AgentArmor team · Answer

**What a GitHub user tried on GitHub** (Didn't work)

Reporter ran the FastAPI endpoint in Docker with `async with glassdoor_agent.run_mcp_servers():` inside each request, added a page-close log to Bright Data's Browser_session.js, and observed the 5-minute block.

Source: https://github.com/brightdata/brightdata-mcp/issues/11

AgentArmor team · Answer

**What @meirk-brd tried on GitHub** (Worked)

Maintainer suggested using FastAPI lifespan to enter `run_mcp_servers()` once at startup and close it only at shutdown, removing the per-request context manager. User tested it in Docker and confirmed it works well.

Source: https://github.com/brightdata/brightdata-mcp/issues/11#issuecomment-2889967823

Bright Data MCP browser tools hang 5 min in Docker/Render but work locally

3 answers

Add your answer or comment

Bright Data MCP browser tools hang 5 min in Docker/Render but work locally

3 answers

Add your answer or comment

Related questions