kagimcp 1.0.0rc2

Kagi MCP server

Kagi MCP server

Tools

The server exposes the following tools, backed by the Kagi API:

• kagi_search_fetch — general web search (workflows: search, news, videos, podcasts, images), with optional inline page extracts, domain include/exclude, date filters, and file-type filter.
• kagi_extract — fetch the full content of a page as markdown.
• kagi_summarizer — summarize any URL (text page, video, audio, etc.) as prose or key takeaways.
• kagi_fastgpt — answer a question with a live web search + LLM synthesis and numbered references.

Setup Intructions

Install uv first.

MacOS/Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Installing via Smithery

Alternatively, you can install Kagi for Claude Desktop via Smithery:

npx -y @smithery/cli install kagimcp --client claude

Setup with OpenAI

Codex CLI

To add the Kagi mcp server to codex cli, you will need to use the following command:

codex mcp add kagi --env KAGI_API_KEY=<YOUR_API_KEY_HERE> -- uvx kagimcp

This will write out the configuration to ~/.codex/config.toml, so if you need to update/rotate your API key, update your key there before running codex again.

Codex CLI comes with its own built-in search (via --search flag), but it is disabled by default. So to deconflict between search and Kagi, just don't enable it.

Setup with Claude

Claude Desktop

// claude_desktop_config.json
// Can find location through:
// Hamburger Menu -> File -> Settings -> Developer -> Edit Config
{
  "mcpServers": {
    "kagi": {
      "command": "uvx",
      "args": ["kagimcp"],
      "env": {
        "KAGI_API_KEY": "YOUR_API_KEY_HERE",
        "KAGI_SUMMARIZER_ENGINE": "YOUR_ENGINE_CHOICE_HERE" // Defaults to "cecil" engine if env var not present
      }
    }
  }
}

Claude Code

Add the Kagi mcp server with the following command (setting summarizer engine optional):

claude mcp add kagi -e KAGI_API_KEY="YOUR_API_KEY_HERE" KAGI_SUMMARIZER_ENGINE="YOUR_ENGINE_CHOICE_HERE" -- uvx kagimcp

Now claude code can use the Kagi mcp server. However, claude code comes with its own web search functionality by default, which may conflict with Kagi. You can disable claude's web search functionality with the following in your claude code settings file (~/.claude/settings.json):

{
  "permissions": {
    "deny": [
      "WebSearch"
    ]
  }
}

Pose query that requires use of a tool

Examples:

• Search: "Who was time's 2024 person of the year?"
• Summarizer: "summarize this video: https://www.youtube.com/watch?v=jNQXAC9IVRw"
• Extract: "extract the full content of https://en.wikipedia.org/wiki/Model_Context_Protocol"
• FastGPT: "what's the latest stable Postgres release and when was it cut?" (returns an answer with citations)

Debugging

Run:

npx @modelcontextprotocol/inspector uvx kagimcp

Local/Dev Setup Instructions

Clone repo

git clone https://github.com/kagisearch/kagimcp.git

Install dependencies

Install uv first.

MacOS/Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Then install MCP server dependencies:

cd kagimcp

# Create virtual environment and activate it
uv venv

source .venv/bin/activate # MacOS/Linux
# OR
.venv/Scripts/activate # Windows

# Install dependencies
uv sync

Setup with Claude Desktop

Using FastMCP CLI

# `pip install fastmcp` if you haven't
fastmcp install claude-desktop /ABSOLUTE/PATH/TO/PARENT/FOLDER/kagimcp/src/kagimcp/server.py --env KAGI_API_KEY=API_KEY_HERE

Manually

# claude_desktop_config.json
# Can find location through:
# Hamburger Menu -> File -> Settings -> Developer -> Edit Config
{
  "mcpServers": {
    "kagi": {
      "command": "uv",
      "args": [
        "--directory",
        "/ABSOLUTE/PATH/TO/PARENT/FOLDER/kagimcp",
        "run",
        "kagimcp"
      ],
      "env": {
        "KAGI_API_KEY": "YOUR_API_KEY_HERE",
        "KAGI_SUMMARIZER_ENGINE": "YOUR_ENGINE_CHOICE_HERE" // Defaults to "cecil" engine if env var not present
      }
    }
  }
}

Pose query that requires use of a tool

Examples:

• Search: "Who was time's 2024 person of the year?"
• Summarizer: "summarize this video: https://www.youtube.com/watch?v=jNQXAC9IVRw"
• Extract: "extract the full content of https://en.wikipedia.org/wiki/Model_Context_Protocol"
• FastGPT: "what's the latest stable Postgres release and when was it cut?" (returns an answer with citations)

Debugging

Run:

# If fastmcp installed (`pip install fastmcp`)
fastmcp dev inspector /ABSOLUTE/PATH/TO/PARENT/FOLDER/kagimcp/src/kagimcp/server.py

# If not
npx @modelcontextprotocol/inspector \
      uv \
      --directory /ABSOLUTE/PATH/TO/PARENT/FOLDER/kagimcp \
      run \
      kagimcp

Then access MCP Inspector at http://localhost:5173. You may need to add your Kagi API key in the environment variables in the inspector under KAGI_API_KEY.

Advanced Configuration

• Level of logging is adjustable through the FASTMCP_LOG_LEVEL environment variable (e.g. FASTMCP_LOG_LEVEL="ERROR")
  • Relevant issue: https://github.com/kagisearch/kagimcp/issues/4
• Summarizer engine can be customized using the KAGI_SUMMARIZER_ENGINE environment variable (e.g. KAGI_SUMMARIZER_ENGINE="daphne")
  • Learn about the different summarization engines here
• Per-tool request timeouts (in seconds) can be set via environment variables. Defaults are tuned for typical latency of each endpoint:
  • KAGI_SEARCH_TIMEOUT — search requests (default: 10)
  • KAGI_EXTRACT_TIMEOUT — page extraction (default: 30)
  • KAGI_SUMMARIZER_TIMEOUT — summarization, which can be slow on long videos/docs (default: 30)
  • KAGI_FASTGPT_TIMEOUT — FastGPT answers (default: 10)
• Transient failures (HTTP 429/500/502/503/504, connection errors, timeouts) are retried with exponential backoff + jitter. Configure max retry attempts via KAGI_MAX_RETRIES (default: 2, i.e. 3 total attempts). Set to 0 to disable retries.
• Tool parameters can be hidden from the LLM via the KAGI_HIDDEN_PARAMS environment variable (comma-separated list). Hidden params fall back to their defaults, reducing context-window noise when you don't need fine-grained control.
  • Hideable params: workflow, extract_count, limit, include_domains, exclude_domains, time_relative, after, before, file_type (search).
  • Example: KAGI_HIDDEN_PARAMS="extract_count,after,before,time_relative,include_domains,exclude_domains" trims the search tool down to query, workflow, limit.
• There may be more secure ways of plugging into the MCP. A user wrote down some details here
• The --http cli option can be used to toggle streamable HTTP transport on. Can use along with --port and --host args.