Hive MCP Server

The MCP surface is intentionally small — two tools for agent integration:

• •search — query workspace state
• •execute — run bounded Python code

Everything durable still goes through the hive CLI and the .hive/ substrate.

Setup

Example MCP config:

{
  "mcpServers": {
    "hive": {
      "command": "hive-mcp",
      "env": {
        "HIVE_BASE_PATH": "/path/to/workspace"
      }
    }
  }
}

HIVE_BASE_PATH should point at the workspace you want the server to search and execute against.

If running from a local checkout instead of an installed package, uv run hive-mcp is the equivalent launch command.

Available Tools

search

Search workspace state, API docs, examples, schemas, and project summaries.

Arguments:

• •query required
• •scopes optional — filter to specific indexes (api, examples, project, workspace)
• •limit optional, defaults to 8

Example:

{
  "name": "search",
  "arguments": {
    "query": "run acceptance",
    "scopes": ["api", "workspace"],
    "limit": 5
  }
}

execute

Run bounded local code with a typed Hive client.

Arguments:

• •code required
• •language optional, Python only for now
• •profile optional
• •timeout_seconds optional

Example:

{
  "name": "execute",
  "arguments": {
    "language": "python",
    "profile": "default",
    "code": "result = hive.task.ready(limit=3)"
  }
}

Important Limits

• •execute is intentionally bounded and time-limited
• •Oversized execute payloads are rejected
• •The Python runner is the only supported language today
• •This MCP surface is not a substitute for the CLI

When To Prefer The CLI

Use the CLI for:

• •Task creation, updates, claims, and releases
• •Run lifecycle actions (start, finish, promote, steer)
• •Evaluator configuration and PROGRAM.md management
• •Projection sync and cache management
• •Campaign and portfolio operations
• •Memory observe, reflect, and search
• •Migration and workspace bootstrap

Use MCP when an agent needs fast search or a local execution sandbox inside a host application.