Kagimcp

What kagisearch/kagimcp MCP server does

The kagisearch/kagimcp MCP server gives MCP-compatible clients access to two Kagi API operations: web search and page extraction. Search results can cover web, news, videos, podcasts, and images. The search tool also supports optional page extracts, search filters, Kagi lenses, and numbered results that users can reference in later instructions.

The extraction tool retrieves the full content of a web page and returns it as Markdown. This is useful when a search result or user-provided URL needs to be read in detail rather than summarized from snippets.

The project no longer exposes the earlier kagi_fastgpt or kagi_summarizer tools. The README says those tools may return in a future release.

How it works

Local deployments communicate with an MCP client over stdio. The recommended package runner is uvx, which starts the published kagimcp package. Requests use the Kagi API and require a Kagi API key supplied through KAGI_API_KEY.

The project also provides a hosted MCP endpoint at https://mcp.kagi.com/mcp. HTTP-capable clients can connect to that endpoint and send the key as a Bearer token in the Authorization header. OAuth2 is not supported at present. For self-hosting, the server can run in streamable HTTP mode, and its HTTP mode accepts a Bearer key per request rather than relying on one server-wide key.

Setup and configuration

For a local client configuration, install uv and run uvx kagimcp with KAGI_API_KEY in the environment. Claude Desktop configuration uses a local command with uvx and the package name as its argument. The repository also documents setup for Codex CLI, Claude Code, Kiro, OpenCode, and Smithery.

KAGI_API_KEY is required. Optional settings include FASTMCP_LOG_LEVEL, search and extraction timeouts, KAGI_MAX_RETRIES, and KAGI_HIDDEN_PARAMS. The hidden-parameter setting can remove selected search controls from the schema shown to the language model, including domain filters, date ranges, file type, result limits, and lens selection.

Tools and capabilities

• kagi_search_fetch: Runs Kagi searches across web, news, videos, podcasts, and images. It supports optional extracts, filters, lenses, and numbered results.
• kagi_extract: Fetches a web page’s full content as Markdown.

The kagisearch/kagimcp MCP server can run over local stdio or streamable HTTP. A Dockerfile is included for HTTP deployment, and the container respects the PORT environment variable used by several hosting platforms.

Limitations and notes

A Kagi API key is needed for local use and hosted requests. The hosted endpoint does not currently support OAuth2. Search and extraction have separate timeout settings, and retries can be disabled by setting KAGI_MAX_RETRIES to 0.

The kagisearch/kagimcp MCP server is focused on Kagi search and extraction. It does not currently include the removed FastGPT or summarizer tools. The repository documents a pinned prerelease installation form when a specific prerelease version is required.