Anyfetch · comparison
Web search MCP without an API key: the free options, tested
A web search MCP server without an API key lets your agent search today, with no account and no monthly limit to watch. This page lists the keyless options, what each one searches, where it runs, and how well it reads the pages it finds.
Direct answer
You can give an AI agent web search through MCP with no API key in five ways: Anyfetch, the DuckDuckGo MCP server, a SearXNG server that you host, and the keyless use of the hosted Exa and Firecrawl servers. Anyfetch searches Brave, Yahoo, and DuckDuckGo from your own machine and reads the result pages in the same call. A search is only useful when the agent can read what it finds: in our test of 97 popular sites that block AI agents, Anyfetch read 82, and the DuckDuckGo MCP server read 49.
The keyless options
| Server | What it searches | Where it runs | Limits | Blocked sites read |
|---|---|---|---|---|
| Anyfetch | Brave, then Yahoo, then DuckDuckGo. Also GitHub, Stack Overflow, Hacker News, arXiv, and research papers. | On your machine | None of its own. A search engine can slow down one address. | 82 of 97 |
| DuckDuckGo MCP server | DuckDuckGo | On your machine | 30 searches and 20 page reads a minute by default | 49 of 97 |
| SearXNG with an MCP server | The engines that you turn on in SearXNG | On a SearXNG server that you host | You run and update the server. Engines can block its address. | Not measured |
| Exa MCP, with no key | The Exa index | In the Exa cloud | A rate limit for use with no key | 78 of 97 |
| Firecrawl MCP, with no key | Firecrawl search | In the Firecrawl cloud | A rate limit for use with no key. Other tools need a key. | Not measured. The rate limit was reached on our server after a few pages. |
Two more servers come up in every list, and both need a key. The Brave Search MCP server needs a Brave API key, and Brave asks for a credit card even for its free credits. The Tavily MCP server needs a Tavily API key.
What your agent already has
| Harness | Built-in search | What to know |
|---|---|---|
| Claude Code | WebSearch | It uses the search of Anthropic. You cannot change the engine, and it is not available on every provider. |
| Codex CLI | A hosted web_search tool | Local sessions use a cached index by default. Live results need --search or a setting. |
| Gemini CLI | google_web_search | It returns a summary with citations, not a list of results. |
| Hermes Agent | web_search | With no key, it rotates between the free use of five hosted services. |
| Cline | None | It has a tool that fetches a page, but no search tool. |
A built-in search is enough when the answer is in the result text. It stops when the agent must open the result and the site answers 403. A search MCP server that also reads pages closes that gap.
Search is only half: can it read the result?
- 82 of 97
- 78 of 97
- 49 of 97
We read each of 97 popular sites once with each tool, on 2026-10-12, from one cloud server with no proxy. A page counts as read when the tool returns at least 500 characters of text and the text is not a bot check. Most of the sites refuse AI agents in some way. A few open sites, such as GitHub and Wikipedia, are in the list as a control. We make Anyfetch and we ran this test, so the list of sites, the script, and the result for every site are public. Run it yourself.
Exa reads many of these sites because it answers from its own stored copy of a page. In our test, each answer stopped at about 3,000 characters, and its copy of the New York Times home page was 26 days old. Anyfetch and the DuckDuckGo MCP server ask the site itself, from your machine, at the time of the call.
We also asked 10 research questions that give no URL, so the agent must search first and then read. A Claude Sonnet agent with only Anyfetch answered 10 of 10 correctly in both of two runs. The same agent with the built-in WebSearch and WebFetch of Claude Code also answered 10 of 10. On 13 questions about a page on a site that blocks agents, the result was 12 against 7.
Set up keyless search with Anyfetch
curl -fsSL https://aliasmode.github.io/anyfetch/install.sh | shirm https://aliasmode.github.io/anyfetch/install.ps1 | iexThe installer connects Anyfetch to Claude Code, Codex, Gemini CLI, Cursor, Hermes Agent, and the other agents it finds. With Node.js, any MCP client can start it with no install:
{
"mcpServers": {
"anyfetch": { "command": "npx", "args": ["-y", "@aliasmode/anyfetch", "mcp"] }
}
}| Input | Meaning |
|---|---|
| query | The search query. Quoted phrases work. |
| site | Keep results from one domain, for example docs.rs. |
| time | day, week, month, or year. |
| read | Also return, from this many top results, the parts of each page that match the query. |
| provider | auto, brave, yahoo, duckduckgo, searxng, github, stackoverflow, hackernews, arxiv, or papers. |
anyfetch search "postgres vacuum freeze" --read 3
anyfetch search "tokio select" --site docs.rs --time year
anyfetch search "retrieval augmented generation" --provider papersDo you run your own SearXNG? Anyfetch can use it: set provider to searxng and give the URL of your server.
When to pay for a search API
- You run many agents from one address. A search engine slows down an address that asks too often, and a paid API has a known limit.
- You need an agreement with the provider about uptime and use.
- You need a search made for one job, such as semantic search over a private index.
For one developer with a coding agent, or a small team, keyless search is enough. When your agents run on a server, a residential IP gives them a home address. NobleProxy is our own service for that.
Frequently asked questions
Is there a web search MCP server that needs no API key?
Yes. Anyfetch and the DuckDuckGo MCP server run on your machine and need no key. A SearXNG server that you host needs no key. The hosted Exa and Firecrawl MCP servers allow limited use with no key.
Does the Brave Search MCP server need an API key?
Yes. It needs a Brave Search API key, and Brave asks for a credit card even for the free monthly credits. Anyfetch searches Brave with no key.
How do I add web search to Claude Code with no API key?
Claude Code has a built-in WebSearch. To add a search that also reads blocked pages, run: claude mcp add --scope user anyfetch -- npx -y @aliasmode/anyfetch mcp.
Is keyless web search rate limited?
Anyfetch has no limit of its own. The search engines can slow down one address that sends many searches. Anyfetch then uses the next engine: Brave, then Yahoo, then DuckDuckGo.
Can a keyless search MCP server also read the pages it finds?
Some can. In our test of 97 popular sites that block AI agents, Anyfetch read 82 and the DuckDuckGo MCP server read 49.