WEB SEARCH / Kagi Search
Kagi MCP Server
Search with Kagi and extract selected pages through a hosted or local MCP server.
“It exposes search and extraction tools to MCP-compatible clients.”
01 / THE REASONING
Why this made the selection.
- Publisher-maintained integration with a documented hosted or local setup and explicit data/cost boundaries.
- Provides both hosted bearer access and a documented per-request identity model for self-hosting.
- The pinned README is Kagi search and page extraction, as a hosted MCP URL or a local uvx process.
- OAuth is not implemented on this pin, and the older FastGPT and summarizer tools are gone.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- README.md
- LICENSE
- pyproject.toml
- Round 2 reread of acquired publisher setup and permission evidence; see immutable wave review.
Our findings
- README and pyproject.toml distinguish hosted bearer authentication, local runtime and separate search/extract timeouts.
- The supplied setup uses the hosted endpoint and does not execute or install the Python package.
- The server exposes kagi_search_fetch and kagi_extract. The README says kagi_fastgpt and kagi_summarizer have been removed.
- The hosted URL is https://mcp.kagi.com/mcp. OAuth2 is not supported yet, so the client sends a Kagi API key as a bearer token.
- Self-hosted HTTP mode is multi-tenant: each request supplies its own key instead of one server-wide environment variable. A local uvx run still needs KAGI_API_KEY.
- KAGI_HIDDEN_PARAMS can remove selected search arguments from the model-facing schema, and retry/timeout settings are separately configurable.
Not established by this review
- Server startup, authentication, live tool execution and security isolation were not tested.
- Published package/container parity with a source revision and hosted implementation internals were not tested.
- The server was not started and its tools were not called.
- A pin does not verify transitive dependencies.
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Connect a server deliberately.
DOCUMENTED COMMAND
claude mcp add kagi -e KAGI_API_KEY="YOUR_API_KEY_HERE" -- uvx kagimcp Copying does not execute this command. It may retrieve a newer version than the reviewed source.
- The command above is the README's Claude Code registration; replace the placeholder with your key. It was not run.
- Other clients use the README's uvx kagimcp block, or the hosted https://mcp.kagi.com/mcp with an Authorization bearer header. Do not commit the key.
Before you start
- uv
- Python 3.12
- A Kagi API key
Documented configuration example
Use the client and authentication context described in the setup guidance. Configuration formats vary between hosts.
{
"mcpServers": {
"kagi": {
"command": "uvx",
"args": [
"kagimcp"
],
"env": {
"KAGI_API_KEY": "YOUR_API_KEY_HERE"
}
}
}
} 04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
https://github.com/kagisearch/kagimcp/blob/55b38d20c67f1406f2c284af776de395297a75cc/README.md
Supports: summary, install, configExample, access, limitations, bestFor, review, compatibility, transports
https://github.com/kagisearch/kagimcp/blob/55b38d20c67f1406f2c284af776de395297a75cc/pyproject.toml
Supports: summary, install, configExample, access, limitations, bestFor, review, compatibility, transports
https://github.com/kagisearch/kagimcp/blob/55b38d20c67f1406f2c284af776de395297a75cc/LICENSE
Supports: license