annotators-mcp

An MCP server for annotators.ai. It lets an agent ask for a human check of its own output, have items labeled, and run quick surveys, from Claude Code, Claude Desktop or any other MCP client.

Tools

Before you start

1. Sign up at annotators.ai/console, create a workspace, and make a key under API keys.
2. Create a task or two. A task says what an item looks like and which fields to fill. Tasks are created with the API or the Python SDK; the agent finds them with list_tasks.
3. Install uv, which runs the server with uvx annotators-mcp without a separate install step. Or pip install annotators-mcp and use the annotators-mcp command in place of uvx annotators-mcp below.

Claude Code

claude mcp add annotators -e ANNOTATORS_API_KEY=ak_... -- uvx annotators-mcp

Add -s user to have it in every project, not only the current one.

Claude Desktop

Open Settings → Developer → Edit Config (that's claude_desktop_config.json, in ~/Library/Application Support/Claude/ on macOS and %APPDATA%\Claude\ on Windows) and add:

{
  "mcpServers": {
    "annotators": {
      "command": "uvx",
      "args": ["annotators-mcp"],
      "env": { "ANNOTATORS_API_KEY": "ak_..." }
    }
  }
}

Restart Claude Desktop. If it can't start the server, put the full path from which uvx in command; apps started from the Dock don't always see your shell's PATH.

Other MCP clients take the same command, arguments and environment.

Who reviews

The tools leave workforce out unless the agent passes one, so items go to your workspace's default: your own team (private), meaning the members of your console workspace and the reviewers you invite. You review under Review in the console.

The annotators.org crowd (open, vetted) is on a waitlist. Join it under Review in the console. Until your workspace has crowd access, a request for the crowd comes back as an error that says so.

Settings

Links

• annotators.ai
• API reference
• Python SDK