Baidu MCP

Baidu MCP

A Model Context Protocol server for Baidu web search and page extraction.

Fork of Evilran/baidu-mcp-server, modernised for the current MCP SDK: stdio and streamable HTTP, structured tool output, and a browser backend that can reuse the browser your IDE or agent harness already runs (over CDP) instead of downloading its own Chromium.

What changed from upstream

Area	Upstream	This fork
SDK	mcp>=1.3 (breaks on mcp 2.x: mcp.server.fastmcp was removed)	mcp>=2,<3 (MCPServer)
Transport	stdio only	stdio (default) and --transport streamable-http
Browser	forces a Playwright Chromium download; crashes if absent	--browser auto|cdp|playwright|http, degrades to plain HTTP
IDE / harness browser	unsupported	attaches over CDP: --cdp-url, or auto-discovers the oh-my-pi browser
Result URLs	m.baidu.com/... tracking redirects	the real target URL (data-log.mu)
Titles	often None	parsed from [data-module="title"], widget cards filtered out
Content fetch	150-character teaser, buried in deep_mode	fetch_url tool, readable Markdown up to max_chars
Tool output	prose string	structured JSON (outputSchema), readOnlyHint annotations
Prompt-injection surface	—	bot-challenge pages detected and reported as errors

Install

# no browser download, works out of the box (plain HTTP transport)
uvx baidu-mcp-server

# with browser support (CDP attach or bundled Chromium)
uvx --from "baidu-mcp-server[browser]" baidu-mcp-server

pip install baidu-mcp-server / pip install "baidu-mcp-server[browser]" work too.

MCP client configuration

stdio (default)

{
  "mcpServers": {
    "baidu": {
      "command": "uvx",
      "args": ["--from", "baidu-mcp-server[browser]", "baidu-mcp-server"]
    }
  }
}

Streamable HTTP

baidu-mcp-server --transport streamable-http --host 127.0.0.1 --port 8000
# endpoint: http://127.0.0.1:8000/mcp

Browser backends

Baidu increasingly answers automated requests with a captcha. A real browser with a real fingerprint is the reliable path, and you may already have one running.

--browser	Behaviour
auto (default)	Prefers a harness/CDP browser, and switches to it as soon as one appears; else a launched Chromium, else plain HTTP
cdp	Attach to a running browser — no Chromium download, shares its network/TLS fingerprint. Waits up to --cdp-wait seconds for one to appear, then errors
playwright	Launch a local Playwright Chromium (needs playwright install chromium)
http	No browser at all: curl_cffi (Chrome TLS impersonation), httpx fallback

Reusing the oh-my-pi / IDE browser

# explicit endpoint
baidu-mcp-server --browser cdp --cdp-url http://127.0.0.1:9222

# or just let it auto-discover the oh-my-pi managed Chromium
baidu-mcp-server --browser auto

Discovery is re-run every time a browser is (re)connected, so an auto backend that had to launch its own Chromium moves over to the harness browser within ~30 s of it starting, and reattaches if the harness browser restarts. auto/cdp resolve the endpoint in this order:

1. --cdp-url
2. BAIDU_MCP_CDP_URL
3. BROWSER_CDP_URL
4. oh-my-pi's running browser: ~/.omp/run/daemons/*/omp.browser*.profile/DevToolsActivePort

Any Chrome started with --remote-debugging-port=9222 works, as do CDP proxies from IDEs and agent harnesses. The server creates an isolated browser context (own cookies, mobile UA), so it does not touch your open tabs.

Tools

search

{
  "query": "python mcp server",
  "max_results": 6,
  "deep_mode": false
}

Returns:

{
  "query": "python mcp server",
  "engine": "baidu",
  "transport": "cdp",
  "results": [
    {
      "title": "mcp 1.8.0 - mcp · PyPI",
      "url": "https://pypi.org/project/mcp/1.8.0/",
      "snippet": "We recommend using uv to manage your Python projects...",
      "position": 1
    }
  ]
}

deep_mode: true additionally fetches each result page and appends a short content preview.

fetch_url

{ "url": "https://example.com/article", "max_chars": 8000 }

Returns { "url", "content", "truncated" }, where content is readability-extracted Markdown. Client-side redirects (window.location.replace, <meta refresh>) are followed.

Rate limiting

Search: 30 requests/minute. Content fetching: 20 requests/minute. Excess requests wait instead of failing.

Configuration

Flag	Env var	Default
--transport	BAIDU_MCP_TRANSPORT	stdio
--host	BAIDU_MCP_HOST	127.0.0.1
--port	BAIDU_MCP_PORT	8000
--browser	BAIDU_MCP_BROWSER	auto
--cdp-url	BAIDU_MCP_CDP_URL	auto-discovered
--cdp-wait	BAIDU_MCP_CDP_WAIT	0 (seconds to wait for a CDP endpoint in cdp mode)
--log-level	BAIDU_MCP_LOG_LEVEL	INFO

Smoke test without an MCP client:

baidu-mcp-server --query "南京天气" --browser http

Requirements

Python 3.10+. Works without any browser download; [browser] adds Playwright for the CDP and launched-Chromium backends.

License

MIT. Original work © Evilran; this fork © ZanDev32.