What specific tools does the Kagi MCP Server expose?

It exposes two main tools: kagi_search_fetch for web, news, video, and image searches with optional extraction, and kagi_extract for pulling full URL text as clean Markdown.

How can I deploy the Kagi MCP Server?

You can run it locally using the uvx kagimcp command or a Docker image, or use the hosted remote MCP endpoint at mcp.kagi.com with Bearer-token authentication.

Is the Kagi MCP Server free to use?

No, it requires a paid Kagi API key with pay-per-use billing. While the server code is MIT-licensed, the underlying search service is not free.

Which MCP-compatible clients are supported?

The server supports clients including Claude Desktop, Claude Code, Codex CLI, Kiro, OpenCode, and Smithery, with documented setup recipes for each.

Are the FastGPT and Universal Summarizer tools currently available?

No, these tools were part of an earlier release and are currently absent during a rewrite, though they are on the roadmap to return.

Kagi MCP Server

Official Kagi Search MCP server: search and extract the web from any MCP-compatible agent.

Paid· Free: $0 USD · Team: $4 USD per user/month · Enterprise: $21 USD per user/monthMCP Servers

Visit website →

In short

Kagi MCP Server is an official, MIT-licensed tool that allows AI agents to use Kagi's search and content extraction capabilities. It is best for engineers who want a cleaner, ad-free web substrate for their coding assistants or research pipelines.

Best for

Engineers using Claude Code, Claude Desktop, or other MCP clients who already pay for Kagi and want their agent to search and read the web through Kagi's cleaner index instead of a generic SERP API.

Skip if

Hobbyists who want a free web-search MCP, or teams that need summarizer/FastGPT tools today, or shops that require an on-prem search stack with no third-party API dependency.

Kagi MCP Server is the official Model Context Protocol server from Kagi Search that lets any MCP-compatible client (Claude Desktop, Claude Code, Codex CLI, Kiro, OpenCode, Smithery) call Kagi's search and content extraction directly from an agent loop. It exposes two tools: kagi_search_fetch, which runs web, news, video, podcast, and image searches with optional page extraction, filtering, and Kagi lenses; and kagi_extract, which pulls the full text of a URL back as clean Markdown suitable for an LLM context window. FastGPT and Universal Summarizer tools were part of an earlier release, were pulled during the rewrite, and are on the roadmap to return. The project is a small MIT-licensed Python package installable with uvx kagimcp, shippable as a Docker image, or reachable as a hosted remote MCP endpoint at https://mcp.kagi.com/mcp with Bearer-token auth. Configuration is a handful of environment variables (KAGI_API_KEY, log level, timeouts). It is aimed at engineers who want a research-grade, ad-free search substrate for agents and coding assistants without hand-rolling their own SerpAPI or Brave wrapper, and who already trust Kagi's ranking and de-junked results over Google/Bing SERPs. Typical workflows include grounding an agent's answers with fresh citations, letting Claude Code look up library docs or error strings mid-task, deep-research runs that chain search then extract on the top N hits, and building small internal RAG helpers that prefer Kagi's cleaner index over general web scraping.

Editor's take

If you already pay for Kagi, this is the search-and-fetch MCP to wire into Claude Code first. It is small, official, MIT-licensed, and gives an agent a noticeably less spammy view of the web than the usual SERP-wrapper MCPs. The missing FastGPT and summarizer tools are the one wart, and the pay-per-use API key means it is not the pick if you want free.

— The AI Tool Bible editorial team

Pros

✅ Official, first-party MCP server maintained by Kagi (not a community wrapper)
✅ Two clean tools that cover the 90% agent case: search and full-page Markdown extraction
✅ Both hosted (mcp.kagi.com) and local (uvx/Docker) deployment modes
✅ Documented setup recipes for Claude Desktop, Claude Code, Codex CLI, Kiro, OpenCode, Smithery
✅ MIT-licensed Python code, easy to fork or audit
✅ Rides Kagi's ad-free, spam-de-boosted index rather than generic SERP APIs
✅ Supports Kagi lenses and filters so agents can scope searches to news, docs, etc.

Cons

⚠️ Requires a paid Kagi API key with pay-per-use billing; not usable free
⚠️ FastGPT and Universal Summarizer tools are currently absent (rewrite in progress)
⚠️ Python + uv toolchain assumption raises the bar vs. a single-binary MCP server
⚠️ Small surface area: no built-in caching, batching, or spend controls beyond Kagi's invoice cap
⚠️ Hosted endpoint means your queries and Bearer token traverse Kagi infra, not a concern for everyone but worth noting for regulated workloads

Use cases

agent web search groundingClaude Code documentation lookupdeep-research pipelinesURL to Markdown extraction for RAGnews and video search from an LLMKagi lens-scoped queries in agentscitation gathering for research assistantscoding agent error-message lookups

Frequently asked

What specific tools does the Kagi MCP Server expose?: It exposes two main tools: kagi_search_fetch for web, news, video, and image searches with optional extraction, and kagi_extract for pulling full URL text as clean Markdown.
How can I deploy the Kagi MCP Server?: You can run it locally using the uvx kagimcp command or a Docker image, or use the hosted remote MCP endpoint at mcp.kagi.com with Bearer-token authentication.
Is the Kagi MCP Server free to use?: No, it requires a paid Kagi API key with pay-per-use billing. While the server code is MIT-licensed, the underlying search service is not free.
Which MCP-compatible clients are supported?: The server supports clients including Claude Desktop, Claude Code, Codex CLI, Kiro, OpenCode, and Smithery, with documented setup recipes for each.
Are the FastGPT and Universal Summarizer tools currently available?: No, these tools were part of an earlier release and are currently absent during a rewrite, though they are on the roadmap to return.