Rasdaman MCP Server

Rasdaman MCP Server

This tool enables users to interact with rasdaman in a natural language context. By exposing rasdaman functionality as tools via the MCP protocol, an LLM can query the database to answer questions like:

• "What datacubes are available?"
• "What are the dimensions of the 'Sentinel2_10m' coverage?"
• "Create an NDVI image for June 12, 2025."

The MCP server translates these tool calls into actual WCS/WCPS queries that rasdaman can understand and then returns the results to the LLM.

Installation

pip install rasdaman-mcp

Usage

First the connection from the MCP server to rasdaman needs to be configured, either through environment variables:

• RASDAMAN_URL: URL for the rasdaman server
• RASDAMAN_USERNAME: Username for authentication
• RASDAMAN_PASSWORD: Password for authentication

or command-line arguments to the rasdaman-mcp tool:

• --rasdaman-url: URL for the rasdaman server (default RASDAMAN_URL env variable or http://localhost:8080/rasdaman/ows).
• --username: Username for authentication (default RASDAMAN_USERNAME env variable or rasguest).
• --password: Sets the password for authentication (default RASDAMAN_PASSWORD env variable or rasguest).

Then the MCP is ready to be used with an AI agent tool, in one of two modes: stdio (default) or http.

`stdio` Mode

Used for direct integration with clients that take over managing the server process and communicate with it through standard input/output. Generally in your AI tool you need to specify the command to run rasdaman-mcp:

rasdaman-mcp --username rasguest --password rasguest --rasdaman-url "..."

Example for enabling it in gemini-cli:

gemini mcp add rasdaman-mcp "rasdaman-mcp --username rasguest --password rasguest"

Benefits:

• Simplicity: No need to manage a separate server process or ports.
• Seamless Integration: Tools are transparently made available to the LLM within the client environment.

`http` Mode

This mode starts a standalone Web server listening on a specified host/port, e.g:

rasdaman-mcp --transport http --host 127.0.0.1 --port 8000 --rasdaman-url "..."

The MCP server URL to be configured in your AI agent would be http://127.0.0.1:8000/mcp with transport streamable-http. For example, for Mistral Vibe extend the config.toml with a section like this:

[[mcp_servers]]
name = "rasdaman-mcp"
transport = "streamable-http"
url = "http://127.0.0.1:8000/mcp/"

Benefits:

• Scalability: The MCP server can be containerized (e.g., with Docker) and deployed as a separate microservice.
• Decoupling: Any client that can speak HTTP (e.g., curl, Python scripts, web apps, other LLM clients) can interact with the tools.
• Testing: Allows for direct API testing and debugging, independent of an LLM client.

AI agents

Once an AI agent is configured with access to rasdaman-mcp, it becomes capable of using several tools:

• list coverages in the configured rasdaman
• get the details of a particular coverage
• execute processing/analytics queries based on a description in natural language

Examples

The following examples demonstrate the interaction with an AI agent using the rasdaman MCP server.

Listing Coverages

Describing a Coverage

Executing a Query

Query Result Visualization

Natural Language Query Suggestion

Development

Setup

1. Clone the git repository:

   git clone https://github.com/rasdaman/rasdaman-mcp.git
   cd rasdaman-mcp/
   

2. Create a virtual environment (if you don't have one):

   uv venv
   

3. Activate the virtual environment:

   source .venv/bin/activate
   

4. Install from source:

   uv pip install -e .
   

Core Components

• Main Application (main.py): This script initializes the FastMCP application. It handles command-line arguments for transport selection, rasdaman URL, username, and password. It then instantiates the RasdamanActions class and decorates its methods to expose them as tools.
• RasdamanActions Class (rasdaman_actions.py): Encapsulates all interaction with the rasdaman WCS/WCPS endpoints. It is initialized with the server URL and credentials, and its methods contain the logic for listing coverages, describing them, and executing queries.
• WCPS crash course (wcps_crash_course.py): A short summary of the syntax of WCPS