Label Studio MCP Server

What HumanSignal/label-studio-mcp-server MCP server does

The HumanSignal/label-studio-mcp-server MCP server exposes Label Studio operations through MCP. It is intended for workflows where an MCP client needs to create or inspect labeling projects, load task data, read annotations, or attach model predictions.

Supported project operations include listing projects, retrieving project details, reading a project's XML labeling configuration, creating projects, and updating existing configurations. Project creation accepts a title and XML label configuration, along with optional settings, and returns project information including a URL.

Task operations cover importing tasks from a JSON file, listing task IDs in a project, retrieving the data payload for an individual task, and fetching that task's annotations. The task-listing operation returns up to 100 task IDs. Prediction creation accepts a task ID and a Label Studio-formatted list of result dictionaries, with optional model version and score values.

How it works

The HumanSignal/label-studio-mcp-server MCP server communicates with Label Studio through the label-studio-sdk. It does not provide a hosted Label Studio instance; the target instance must already be running and reachable from the machine that runs the MCP process.

An MCP client sends structured tool calls, or requests the operations through natural-language interactions supported by that client. A typical workflow is to create a project, prepare a JSON file containing task objects, import the file, inspect task IDs and data, and then submit predictions for selected tasks. Label Studio remains the system of record for projects, tasks, annotations, and predictions.

Setup and configuration

Create or obtain an API key from the Label Studio user account settings. Configure these environment variables for the MCP process:

• LABEL_STUDIO_API_KEY: API key used to authenticate with Label Studio.
• LABEL_STUDIO_URL: Base URL of the Label Studio instance, such as a locally hosted instance at port 8080.

The README shows a uvx configuration that loads the server directly from the GitHub repository and starts the mcp-label-studio entry point. It provides configuration examples for Claude Desktop and Cursor MCP settings. The server must be able to reach the configured Label Studio URL, and imported task files must be available to the process at the path supplied to the import tool.

Tools and capabilities

The HumanSignal/label-studio-mcp-server MCP server includes tools for:

• Listing projects with IDs, titles, and task counts.
• Getting project details and XML labeling configurations.
• Creating projects and changing their labeling configurations.
• Importing task objects from JSON files.
• Listing project task IDs, up to the documented limit of 100.
• Retrieving task data and existing annotations.
• Creating predictions with Label Studio result dictionaries, model-version metadata, and scores.

Prediction result structures must follow Label Studio's expected format. The server uses the official SDK rather than defining a separate data API.

Limitations and notes

A running Label Studio deployment and a valid API key are prerequisites. The server does not describe a built-in authentication flow, so credentials must be supplied through the process environment or the MCP client's server configuration. Task import expects a JSON file containing a list of task objects. The README documents project and task management capabilities, but does not establish support for other Label Studio features beyond those operations.

The documented client examples are Claude Desktop and Cursor. The repository's support path is GitHub Issues, and no license information is provided in the supplied material.