This component provides a Model Context Protocol (MCP) server for building profiles projects through AI applications like Claude in Cursor IDE.

• Requires Python 3.10.x to be installed
• Access to Snowflake with appropriate permissions (read access to input tables for profiles, and write access to a schema where profiles can generate outputs)
• Important: Snowflake MFA (Multi-Factor Authentication) is not supported. If your Snowflake account has MFA enabled, you must use key-pair authentication instead. Key-pair authentication is Snowflake's recommended method for uninterrupted programmatic access and provides security for automated tools.
• Cursor IDE v0.47.x or higher installed for building profiles projects. A free version works in theory but the experience is significantly better with a paid version
• A Rudderstack Personal Access Token

1. Set up the mcp server:

# Run the setup script to install dependencies, validate your environment, and download embeddings:
./setup.sh

This will:

• Check for Python and uv
• Install uv if it is not found
• Create .env if missing
• Install dependencies
• Download and extract embeddings to src/data/
• Update MCP configuration

2. Restart Cursor to apply changes

Once configured, you can use natural language to build a profiles project through chat interface in AI clients such as cursor:

-- Example queries
- build a rudderstack profiles project to calcualte the churn propensity score for the data in snowflake under db RUDDERSTACK_TEST_DB and schema predictions_dev_project
- build a rudderstack profiles project to find revenue metrics for each user in snowflake under db RUDDERSTACK_TEST_DB and schema predictions_dev_project

While the chat experience will work with most LLMs, we recommend using claude 4.0 class of models (ex: sonnet-4)

In Cursor MCP settings, you should see the profiles mcp tool active, with a green button indicating the MCP tools are available. See the below image for reference:

If you don't see the tools, or the profiles mcp server shows as inactive, run this command on terminal within the profiles-mcp directory. This should launch the mcp server directly, and if there are any errors in running the script, this should catch these.

> scripts/start.sh

Then check the log file, profiles-mcp.log within the profiles-mcp directory.

The MCP Server provides following categories of tools:

• Resources tool - about_profiles. It provides the AI agent with static text, which has the required info about the pb concepts, models, syntax etc.
• RAG tools. Example search_profiles_docs. For more open-ended questions that can be answered from our docs, these tools provide a way for the agent to ask specific questions and get answers. These tools are useful for debugging when the agent faces any errors
• Query tools. Example are run_query, input_table_suggestions etc. These use a snowflake connector and can be used to run queries directly on the warehouse
• Profiles utility tools. Example are get_profiles_output_details, get_existing_connections etc, which look at the yaml files, output files, siteconfig etc and provide required context to the AI agent

1. A snowflake connector that connects to the warehouse using credentials provided while setting up the tool. It needs to be to the same account where the project will eventually be built. This need not be in the same database or schema where profiles inputs or outputs will be generated
2. An in-memory Qdrant vector db enables the RAG workflow

1. Integration to Claude desktop automatically
2. More analysis tools on the profiles output tables

For developers working with RAG-related tools, refer to the backend API service at: https://github.com/rudderlabs/profiles-mcp-service

This repository contains the backend implementation that powers the RAG functionality used by tools like search_profiles_docs.

1. Update your Cursor MCP configuration in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "profiles": {
      "command": "/path/to/scripts/start.sh",
      "args": []
    }
  }
}

2. Restart Cursor to apply changes

This project is licensed under the Elastic License v2.0 - see the LICENSE file for details.