Okareo MCP server

Okareo MCP Server

The Okareo MCP server exposes Okareo's evaluation capabilities as MCP tools, allowing AI coding assistants to create and manage scenarios, register models, run evaluations, and execute multi-turn simulations directly from your editor.

For detailed documentation, see the Okareo MCP docs.

Prerequisites

• An Okareo account at app.okareo.com
• A copilot that supports MCP servers (Claude Code, Cursor, or VS Code)
• Python 3.10–3.12 (only for the local install modes; not needed for remote)

Remote MCP (hosted) — recommended

The fastest way to get started. No local install, no Python, no API key in .mcp.json. Browser sign-in handles auth on first connect.

Recommended (OAuth — Claude Code, Claude Desktop, Cursor, VS Code 1.101+)

Add to your copilot's MCP config (typically .mcp.json):

{
  "mcpServers": {
    "okareo": {
      "type": "http",
      "url": "https://tools.okareo.com/mcp"
    }
  }
}

Reload the copilot. It will open a browser to Okareo sign-in once; thereafter the copilot stores the token itself. Your .mcp.json contains no secrets.

Fallback (Bearer header — older clients or headless / CI)

For clients that haven't shipped the MCP OAuth flow yet, paste your API key as a bearer header. Prefer the env-var form over an inline literal:

{
  "mcpServers": {
    "okareo": {
      "type": "http",
      "url": "https://tools.okareo.com/mcp",
      "headers": {
        "Authorization": "Bearer ${env:OKAREO_API_KEY}"
      }
    }
  }
}

Working across multiple Okareo organizations

If you have access to more than one Okareo account, simply Clear Authentication and then re-authenticate with the account you want to work in. You can always ask which account your are in and have acccess to.

Local install (alternative)

Run the MCP server on your own machine. Useful for offline / airgapped environments and for development.

Prefer the Remote MCP section above unless you have a specific reason to install locally (airgapped, custom build, development on this repo). The remote endpoint requires no install and stays current automatically.

For multi-org users, the remote endpoint also exposes tenant management tools (listtenants, switchtenant).

Step 1: Set Your API Key

export OKAREO_API_KEY="your-api-key"

Add this to your ~/.zshrc or ~/.bash_profile for persistence.

Step 2: Configure Your Copilot

Claude Code

Add to .mcp.json:

{
  "mcpServers": {
    "okareo": {
      "command": "uvx",
      "args": ["okareo-mcp"],
      "env": {
        "OKAREO_API_KEY": "${OKAREO_API_KEY}"
      }
    }
  }
}

No pre-install needed — uvx handles it automatically.

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "okareo": {
      "command": "uvx",
      "args": ["okareo-mcp"]
    }
  }
}

Cursor inherits OKAREOAPIKEY from your shell environment.

Alternative: pip install

If you don't have uv installed:

pip install okareo-mcp

Then use "command": "okareo-mcp" instead of "command": "uvx" with "args": ["okareo-mcp"].

Configuration Reference

Available Tools

Scenarios

Shortened. The full README is on GitHub.