free-search-mcp

https://github.com/user-attachments/assets/354b522a-ef6b-4547-bb07-c80c6268a5e8

_{Music: “Ethereal Space (cdk Mix)” by cdk, ccmixter.org/files/cdk/34151, licensed CC BY 3.0.}

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

	Multi-engine	No API key	Smart fallback	PDF/DOCX	FTS5 cache	Filters	Trafilatura	LLM-tuned
nickclyde/duckduckgo-mcp-server	✗	✓	✗	✗	✗	✗	✗	~
mrkrsl/web-search-mcp	✓	✓	✓	✗	✗	✗	✗	~
Aas-ee/open-webSearch	✓	✓	~	✗	✗	✗	✗	~
VincentKaufmann/noapi-google-search-mcp	✗	✓	✓	✓	✓	✗	✗	~
free-search-mcp	✓	✓	✓	✓	✓	✓	✓	✓

"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)

Tool	Description
search(query, ...filters)	Parallel multi-engine search, RRF-merged, title-fuzzy + host-canonical deduped, with optional extractive lead_snippet
research(question, depth?, ...filters)	One-shot: search + fetch top N + return Markdown brief
paper_graph(paper, direction?, limit?)	Walk one paper's citation graph: references, citing works ranked by influence, and Crossref retraction/correction notices. Takes a DOI, an OpenAlex ID, an exact title, or an arXiv reference (arXiv:1706.03762, an arxiv.org/abs/… URL, a 10.48550/arXiv.… DOI, or a bare id such as 1706.03762, 2401.12345v3 or hep-th/9901001 when it is the whole input)
compare(question, urls=[2..5])	Concurrent fetch of 2-5 URLs, side-by-side excerpts keyed by question
fetch(url, render?, inline?, ...)	Fetch any resource: reader-mode Markdown for pages, parsed text for documents, or a description (type/size/dimensions/sha256) for images and binaries. inline=True returns the image itself for vision models
fetch_batch(urls, ...)	Concurrent multi-URL fetch (max 20 per call)
read_doc(source, start?, length?, ...)	Parse PDF / DOCX / XLSX / PPTX / EPUB / CSV / code / zip-tar / HTML / TXT / MD with pagination
extract_structured(url, ...)	Pull JSON-LD / OpenGraph / Twitter cards / microdata via extruct. Long prose fields (articleBody and the like) are clipped, and the result says so in trimmed. This tool returns metadata; use fetch for the text
cache_search(query, limit?, ...)	FTS5 search across previously fetched pages
engines(group?)	The source tree (group, then sub-group, then engine), one line each. group is an enum of the groups that own engines, so a typo fails schema validation
download(url, ...)	Save a file to ${SEARCH_MCP_CACHE_DIR}/downloads by default; files auto-delete after 24h. Set SEARCH_MCP_DOWNLOAD_ENABLED=false to disable it.

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:

Param	Values	Effect
freshness	day / week / month / year	Only results from the last N. Results with no date are kept, and freshness_note says when that is most of them. day / week also add googlenews
include_domains	["python.org", "djangoproject.com"]	Restrict to these domains
exclude_domains	["pinterest.com"]	Remove these
category	a group (paper, finance, news, software, security, weather, …) or a sub-group (paper.biomed, finance.filings, finance.fx, software.python, dataset.ml, …)	Routes to the sources that natively index that kind of thing; the enum in the tool schema lists every value
include_text	"async"	Substring required in title/snippet
exclude_text	"beginner"	Substring forbidden
max_age_hours	24	Accept a cached answer only if it is younger than this. Default 7 days; the tightest of this, the freshness window's TTL, the 6-hour news cap and the category's own cap (weather, finance.fx 1 hour; software, security 6 hours) wins

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:

Where	How
Claude Desktop	open free-search-mcp-<version>.mcpb from the latest release
MCP Registry	io.github.sweetcornna/free-search-mcp
Claude Code, without the plugin	claude mcp add search -s user -- uvx free-search-mcp
Codex, without the plugin	codex mcp add search -- uvx free-search-mcp
Any other MCP client	the stdio command uvx free-search-mcp
Docker	docker compose build, then docker compose run --rm search-mcp

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.

	codex	antigravity
Searches with	OpenAI's web search, the one Codex uses	Google Search, run by a Gemini model
Account	a ChatGPT plan that includes Codex	a Google account with Antigravity
Each search counts against	the plan's Codex usage	the account's Antigravity quota
The provider allows it	yes	no, see the warning below
Opens a sign-in page by itself	on first use, over stdio on a desktop	never
Full guide, with 中文速览	docs/CODEX_SEARCH.md	docs/ANTIGRAVITY_SEARCH.md

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)

Var	Default	Meaning
SEARCH_MCP_CODEX_MODEL, SEARCH_MCP_ANTIGRAVITY_MODEL	latest	latest follows the account's model catalogue, so a new model is used once it is listed; a model name pins one
SEARCH_MCP_CODEX_TIMEOUT, SEARCH_MCP_ANTIGRAVITY_TIMEOUT	60	seconds for one search
SEARCH_MCP_CODEX_AUTO_SIGNIN	true	open the sign-in page on first use
SEARCH_MCP_ANTIGRAVITY_CLIENT_ID, SEARCH_MCP_ANTIGRAVITY_CLIENT_SECRET	read from the Antigravity install	for a machine without Antigravity; set both or neither
SEARCH_MCP_PROXY	empty	the sign-in and the searches go through it

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

The error says	Do this
codex not configured / antigravity not configured	no sign-in is stored on this machine; run the sign-in above
port 1455 and 1457 … are in use, or port 51121	another sign-in (codex login, an open sign-in page) holds the port; close it and retry
the usage limit, quota or rate limit was reached	wait until the reset time the error gives; the keyless engines still work
needs Antigravity's OAuth client	install Antigravity, or set the two client variables
HTTP 403 no valid license from antigravity	set SEARCH_MCP_ANTIGRAVITY_VERSION to the installed Antigravity release; it can also mean Google restricted the account
has expired or was revoked	sign in again

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:

Engine	Provider	Key
brave_api	Brave Search API	SEARCH_MCP_BRAVE_API_KEY
serper	Serper	SEARCH_MCP_SERPER_API_KEY
tavily	Tavily	SEARCH_MCP_TAVILY_API_KEY
google_cse	Google Custom Search	SEARCH_MCP_GOOGLE_CSE_API_KEY + SEARCH_MCP_GOOGLE_CSE_CX
github_code	GitHub	SEARCH_MCP_GITHUB_TOKEN

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:

You want	Use	It needs
The agent to search and read by itself	the tools, as installed	nothing
Your host's own subagents to do the lookup	the quick-search agent definition	a host with subagents
The server to dispatch through Claude Code	SEARCH_MCP_AGENT_BACKEND=claude-code	the claude CLI, logged in
The server to dispatch through Codex	SEARCH_MCP_AGENT_BACKEND=codex	the codex CLI, logged in
The server to call a model endpoint you choose	SEARCH_MCP_AGENT_BACKEND=api	an OpenAI-compatible or Anthropic-compatible URL
Your own code to supply the model	ask(question, answer_with=...) in Python	nothing else

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:

Var	Default	Meaning
SEARCH_MCP_PROXY	empty	outbound proxy (http, https or socks5) for the engines, the browser and fetch
SEARCH_MCP_PROXY_ENGINES	empty	proxy only these engines
SEARCH_MCP_DEFAULT_ENGINES	["duckduckgo","bing","anysearch","mojeek"]	the pool a search uses when it names no engines
SEARCH_MCP_REGION	us-en	locale for engines that take one
SEARCH_MCP_CACHE_DIR	~/.cache/search-mcp	search and page cache
SEARCH_MCP_TOOLS	empty	register only these tools

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.