free-search-mcp

free-search-mcp is a local-first Model Context Protocol server that needs no API key. It lets any LLM (Claude, GPT, a local Ollama model, …) search the web, fetch and clean up pages, and read documents, and you never sign up for a search API.

It combines ideas from several open-source MCP servers in one Python package, and adds the output shaping for LLMs and the reliability work that each of them lacked.

research("how does reciprocal rank fusion work", depth=3)
   ↓
# Research brief: how does reciprocal rank fusion work
_engines: duckduckgo, bing, anysearch · sources: 3 · ~3,400 tokens_

## Sources
- [1] Reciprocal rank fusion | Elasticsearch Reference — <https://…>
- [2] Hybrid Search Scoring (RRF) | Microsoft Learn — <https://…>
- [3] RRF explained in 4 mins — Medium — <https://…>

## Documents
…full Markdown bodies of each page, ready for the LLM to read…

That was one tool call. It returned three sources with their full text, and no API key was involved.

Quick start

You need uv and nothing else: no sign-up, no API key, no clone.

In Claude Code, inside a session:

/plugin marketplace add sweetcornna/free-search-mcp
/plugin install free-search@free-search-mcp

In Codex:

codex plugin marketplace add sweetcornna/free-search-mcp
codex plugin add free-search@free-search-mcp

(In Codex, a [mcp_servers.search] entry left in ~/.codex/config.toml by an earlier codex mcp add silently shadows the plugin's server. Remove it with codex mcp remove search.)

The plugin brings the 11 tools, the verified-research skill (how to check a snippet against its page and its date) and, in Claude Code, the free-search:quick-search agent, all pinned to one version. /plugin update free-search moves to the next release.

Optionally, install Chromium once for the browser-rendered engines (brave, startpage, zhihu, …) and JavaScript-heavy pages. Everything else works without it, and a call that needs it returns this command as its error:

uvx --from free-search-mcp playwright install chromium

Claude Desktop, other clients, a source checkout and Docker are under Install.

Why this exists

"LLM-tuned" means Markdown-first output with token estimates, errors that name the next step, and a research() that turns a search and its fetches into one call. docs/HOW_IT_WORKS.md explains the less obvious columns.

Tools (11)

There are also 5 MCP prompts and 2 resource templates (cache://page/{url}, cache://search/{query_hash}). A twelfth tool, ask, appears only when an answer backend is configured (see Delegating a lookup).

search and research take these filters:

Output is Markdown by default, with provenance and a token estimate in the header. format="json" returns structured data.

How a search runs

• A search that names no engines asks a small keyless pool: duckduckgo, bing, anysearch and mojeek. googlenews joins when recency is asked for, and so360 when the query is written in Chinese.
• category= routes to sources that index that kind of thing (papers, filings, datasets, packages, CVEs, …). Record sources such as pypi, nvd or openmeteo also join by themselves when the question is one they can answer directly.
• An engine that serves a CAPTCHA, a login wall or an off-topic page is benched, a reserve takes its seat, and the response says what happened. A proxy (SEARCH_MCP_PROXY) is the fix for IP gating.
• Every result says how old it is and where its date came from, so the agent can tell a lead from a checked fact.
• engines() prints the whole source tree. engines=["so360", "baidu"] runs exactly the engines named.

No engine that a search reaches by itself needs a key. The details, with measurements, are in docs/HOW_IT_WORKS.md; walls and proxies in docs/PROXY_AND_GATES.md; a tour of the tools with examples in docs/USAGE.md.

Install

The plugin in Quick start is the recommended path, because the server is pinned to the plugin's version and updates with it. The other routes run the same server:

A source checkout with Chromium, a smoke test and client registration in one step:

curl -LsSf https://raw.githubusercontent.com/sweetcornna/free-search-mcp/main/scripts/install.sh | bash -s -- --client claude-code

Over HTTP, uvx free-search-mcp --transport streamable-http --port 8000 serves http://127.0.0.1:8000/mcp. That endpoint has no authentication and fetches any URL for whoever reaches it, so keep it on loopback or put an authenticating proxy in front.

The JSON for Claude Desktop, Cursor, Cline, Continue and Zed, the installer's options and the plugin's token cost are in docs/INSTALL.md. Operating rules for agents are in docs/AGENT_USAGE.md.

Search on your own account: codex and antigravity

Two opt-in engines search on an account you sign in with instead of an API key. Neither is in any pool or route: a search reaches one only when the call names it, and nothing changes for anyone who never signs in.

Warning: Google's Antigravity terms forbid using its sign-in from third-party tools and name suspension of the Antigravity and Gemini CLI accounts as the consequence, and Google has suspended accounts for it. Sign in to antigravity only with an account you accept that risk for.

1. Sign in once

Run the sign-in on the machine the server runs on:

uvx --from free-search-mcp search-mcp-login codex         # the ChatGPT page `codex login` shows
uvx --from free-search-mcp search-mcp-login antigravity   # prints the warning, then Google's consent page
uvx --from free-search-mcp search-mcp-login status        # account, plan, token expiry

(uv run search-mcp-login … in a source checkout.) The settings page, search-mcp-admin, has the same Sign in / 登录 buttons. Tokens are stored in ~/.config/search-mcp/oauth/ (0600) and refreshed automatically. The plugin, uvx and the Desktop bundle all read that directory, so a running server picks up a new sign-in without a restart.

• codex can skip this step: the first search that names it opens the sign-in page and finishes once you approve (stdio on a desktop only; SEARCH_MCP_CODEX_AUTO_SIGNIN=false turns it off). A machine already signed in to the Codex CLI can reuse that sign-in read-only with search-mcp-login codex --use-codex-cli.
• antigravity signs in with Antigravity's own OAuth client, which is not shipped in this package. The sign-in reads it from the Antigravity app installed on the machine (checked on macOS). Without Antigravity installed, set SEARCH_MCP_ANTIGRAVITY_CLIENT_ID and SEARCH_MCP_ANTIGRAVITY_CLIENT_SECRET; a machine that has signed in keeps both in ~/.config/search-mcp/oauth/antigravity.json.

2. Name the engine

search("rust 2024 edition changes", engines=["codex"])
research("what changed in python 3.14 asyncio", engines=["antigravity"])
search("rust 2024 edition changes", engines=["codex", "duckduckgo", "bing"])

In Claude Code or Codex, ask for it in words ("search this with the codex engine"). Filters apply as usual, and results are cached like any other search. Measured on 2026-09-26, a codex search took about 3 s and an antigravity search 10 to 20 s.

3. Settings (all optional)

On a server, over SSH, or in Docker

• Sign in with --no-browser (search-mcp-login codex --no-browser, or antigravity --no-browser), open the printed address in any browser, and approve. The browser then fails to load a 127.0.0.1:1455 or localhost:51121 address; paste that whole address into the terminal.
• Over streamable-http, codex never opens a sign-in page by itself, since it would open on the server. Sign in there with the command above first.
• In Docker, point SEARCH_MCP_CONFIG_DIR at a mounted volume so the tokens outlive the container, and run the sign-in inside it with --no-browser. antigravity there needs the two client variables.

search-mcp-login logout codex or logout antigravity deletes the stored tokens. To revoke Antigravity's access itself, remove it under "Third-party apps and services" in the Google account.

When it does not work

Optional: your own API key

You do not need one, and agents should never ask for one. Five engines run only when a call names them and the operator has set their own key:

Set a key as an environment variable or .env line, or on the local settings page:

uv run search-mcp-admin        # local settings / 本地设置 — http://127.0.0.1:8765

(uvx --from free-search-mcp search-mcp-admin for a plugin or uvx install.) The page is bilingual (中英双语), binds to 127.0.0.1 only, applies a saved key without a restart, and never shows a stored value again. The walkthrough per provider is docs/API_KEYS.md.

Delegating a lookup

Delegation keeps page text out of the caller's context and returns a short answer with dated sources. None of it is on by default:

The settings, the agent files and measured timings are in docs/DELEGATION.md.

Configuration

Nothing is required. Settings are SEARCH_MCP_* variables, read from the environment first, then ./.env, then ~/.config/search-mcp/.env (SEARCH_MCP_CONFIG_DIR moves that directory). The ones people change most:

Every setting is listed in docs/CONFIGURATION.md and documented in .env.example.

Development

git clone https://github.com/sweetcornna/free-search-mcp.git && cd free-search-mcp
uv sync && uv run playwright install chromium
uv run pytest -q                              # offline
SEARCH_MCP_TEST_NETWORK=1 uv run pytest -v    # live, hits the real web

Running claude inside the checkout picks up the repo's .mcp.json, which starts the working tree's server. The architecture is in docs/HOW_IT_WORKS.md, and cutting a release in docs/RELEASING.md.

Credits

This project builds on:

• mrkrsl/web-search-mcp: the httpx-then-Playwright fetch strategy and the multi-engine fallback chain
• Aas-ee/open-webSearch: multi-engine breadth (Bing/DDG/Baidu/Brave/Startpage)
• VincentKaufmann/noapi-google-search-mcp: anti-detection patterns (navigator.webdriver, UA, cookies), the SQLite FTS5 cache idea, and multi-format read_document
• nickclyde/duckduckgo-mcp-server: per-engine rate limiting and LLM-friendly content cleanup
• Mojeek, an independent search index that doesn't gate on User-Agent
• Model Context Protocol and the official Python SDK

License

MIT. See LICENSE.