Label Studio MCP Server

Overview

This project provides a Model Context Protocol (MCP) server that allows interaction with a Label Studio instance using the label-studio-sdk. It enables programmatic management of labeling projects, tasks, and predictions via natural language or structured calls from MCP clients. Using this MCP Server, you can make requests like:

• "Create a project in label studio with this data ..."
• "How many tasks are labeled in my RAG review project?"
• "Add predictions for my tasks."
• "Update my labeling template to include a comment box."

Features

• Project Management: Create, update, list, and view details/configurations of Label Studio projects.
• Task Management: Import tasks from files, list tasks within projects, and retrieve task data/annotations.
• Prediction Integration: Add model predictions to specific tasks.
• SDK Integration: Leverages the official label-studio-sdk for communication.

Prerequisites

1. Running Label Studio Instance: You need a running instance of Label Studio accessible from where this MCP server will run.
2. API Key: Obtain an API key from your user account settings in Label Studio.
3. Python Environment: Python 3.x with uv installed for package management is recommended.

Installation

Follow these instructions to install the server.

git clone https://github.com/HumanSignal/label-studio-mcp-server.git 
cd label-studio-mcp-server

# Install dependencies using uv
uv venv
source .venv/bin/activate 
uv sync

Configuration

The MCP server requires the URL and API key for your Label Studio instance. There are two main ways to configure this:

1. Environment Variables: Set the following environment variables in your terminal session before running the server:

   export LABEL_STUDIO_URL='http://localhost:8080' # Replace with your LS URL
   export LABEL_STUDIO_API_KEY='your_actual_api_key_here'
   
   python label-studio-mcp.py

   The Python script reads these using os.getenv().

2. MCP Client Configuration (e.g., Cursor's mcp.json): If launching the server via an MCP client configuration file, you can specify the environment variables directly within the server definition. This is often preferred for client-managed servers.

   Example mcp.json entry:

   {
     "mcpServers": {
       "label-studio": {
           "command": "uv",
           "args": [
               "--directory",
               "/path/to/your/label-studio-mcp-server", // <-- Update this path
               "run",
               "label-studio-mcp.py"
           ],
           "env": {
               "LABEL_STUDIO_API_KEY": "your_actual_api_key_here", // <-- Your API key
               "LABEL_STUDIO_URL": "http://localhost:8080"
           }
       }
     }
   }

   When configured this way, the env block injects the variables into the server process environment, and the script's os.getenv() calls will pick them up.

Install for Claude Desktop

To install the MCP server for Claude Desktop, you can run the following command from the root of this repository:

mcp install -e . -v LABEL_STUDIO_API_KEY=<your_api_key> -v LABEL_STUDIO_URL=<your_label_studio_url> label-studio-mcp.py

Tools

The MCP server exposes the following tools:

Project Management

• get_label_studio_projects_tool(): Lists available projects (ID, title, task count).
• get_label_studio_project_details_tool(project_id: int): Retrieves detailed information for a specific project.
• get_label_studio_project_config_tool(project_id: int): Fetches the XML labeling configuration for a project.
• create_label_studio_project_tool(title: str, label_config: str, ...): Creates a new project with a title, XML config, and optional settings. Returns project details including a URL.
• update_label_studio_project_config_tool(project_id: int, new_label_config: str): Updates the XML labeling configuration for an existing project.

Task Management

• list_label_studio_project_tasks_tool(project_id: int): Lists task IDs within a project (up to 100).
• get_label_studio_task_data_tool(project_id: int, task_id: int): Retrieves the data payload for a specific task.
• get_label_studio_task_annotations_tool(project_id: int, task_id: int): Fetches existing annotations for a specific task.
• import_label_studio_project_tasks_tool(project_id: int, tasks_file_path: str): Imports tasks from a JSON file (containing a list of task objects) into a project. Returns import summary and project URL.

Predictions

• create_label_studio_prediction_tool(task_id: int, result: List[Dict[str, Any]], ...): Creates a prediction for a specific task. Requires the prediction result as a list of dictionaries matching the Label Studio format. Optional model_version and score.

Example Use Case

1. Create a new project using create_label_studio_project_tool.
2. Prepare a JSON file (tasks.json) with task data.
3. Import tasks using import_label_studio_project_tasks_tool, providing the project ID from step 1 and the path to tasks.json.
4. List task IDs using list_label_studio_project_tasks_tool.
5. Get data for a specific task using get_label_studio_task_data_tool.
6. Generate a prediction result structure (list of dicts).
7. Add the prediction using create_label_studio_prediction_tool.

Running the Server

The MCP server can be run directly using Python:

python label-studio-mcp.py [options]

Options:

• --transport {http|stdio}: Specify the communication transport (default: stdio).
• --port PORT: Set the port number for the HTTP transport (default: 3000).
• --host HOST: Set the host address for the HTTP transport (default: 0.0.0.0).

Ensure the required environment variables (LABEL_STUDIO_URL, LABEL_STUDIO_API_KEY) are set before running.

Contact

For questions or support, reach out via GitHub Issues.