地球科学数据访问服务

earthdata-mcp

An MCP (Model Context Protocol) server providing LLM agents with direct access to NASA's Earth science data APIs, currently the Common Metadata Repository (CMR). This integration enables users to agentically discover, verify, and access Earth science datasets through natural language interfaces.

To start querying Earthdata immediately, see Connecting to the Server below, or the user guide on the documentation site for client setup, example queries, and troubleshooting. The public consumer documentation is published at nasa.github.io/earthdata-mcp and maintained in docs/. For a detailed breakdown of how tool inputs and outputs map to the APIs and schemas behind them, see the Parameter Support Reference.

Available Tools

• get_keywords: Discovers official Earthdata scientific vocabulary terms (from NASA KMS) to translate colloquial user inputs (e.g. "rain") into precise search labels (e.g. "PRECIPITATION AMOUNT").
• get_collections: Searches for datasets (collections) using scientific keywords, instruments, platforms, or spatial/temporal constraints.
• get_granules: Searches for specific data files (granules) within a collection. Used to verify actual data availability for a given time and location.
• get_services: Discovers data access endpoints (OPeNDAP, Harmony) and visualization layers (WMS/WMTS) associated with a collection.
• get_tools: Finds web portals (e.g., Giovanni, Worldview) and downloadable software (e.g., Panoply) associated with a collection, returning URLs and deep-linking templates.
• get_citations: Discovers citation records (publications, DOIs) associated with a collection, or looks up citations directly by identifier.
• get_variables: Discovers scientific variables and measurements associated with a collection, or looks up variables by keyword. Use this to understand specific data parameters (scale, offset, fill values) before downloading or analyzing data.

Agent Workflow Instructions

The server provides system instructions (prompts/instructions.py) that enforce a Discover → Verify → Access workflow for LLM clients:

1. Discover: Find relevant collections using get_collections.
2. Verify: Use get_granules to confirm data actually exists for the user's requested region/time, as collections often declare global coverage regardless of gaps.
3. Access: Instruct users to use the earthaccess Python library for authentication and downloading, providing relevant code snippets.

Project Structure

The repository is structured around a few core domains:

• server.py & loader.py: The FastMCP server entry point and dynamic tool registration logic.
• prompts/: System prompts and instructions that define the LLM's workflow and persona.
• tools/: Self-contained MCP tools wrapping NASA CMR APIs (get_collections, get_granules, get_keywords, get_services, get_tools, get_citations, get_variables).
• models/: Pydantic models for tool input validation and standardized CMR API responses.
• tests/: Comprehensive test suite (using pytest) covering server initialization, tool logic, and mocked CMR API responses.
• docs/: Project documentation separated into consumers/ and developers/.

Note on Legacy Code: The ingestion and embedding pipelines (including the discover_data tool, lambdas/ directory, and associated infrastructure) are currently being deprecated. The architecture is transitioning to rely purely on direct, real-time CMR API integrations.

For Consumers: Connecting to the Server

The Earthdata MCP server is deployed remotely and communicates via the official Model Context Protocol Streamable HTTP transport. To use the server, configure your MCP-compatible client to point to our endpoint.

Connection URL

Configure your client to connect to the following HTTP endpoint. Most MCP clients require these standard parameters:

• Transport Type: streamable-http
• URL: https://cmr.earthdata.nasa.gov/mcp/v1
• Timeout: 60000 (Recommended: 60 seconds to allow for complex spatial/temporal queries)

Works with:

• Claude Code CLI
• VS Code MCP extensions
• LibreChat
• Any MCP-compatible client that supports Streamable HTTP transport

For Developers: Local Environment

If you want to contribute to the server or run it locally, follow these steps.

Prerequisites

• Python 3.13+
• uv package manager

Installation

# Clone and enter directory
git clone <repo-url>
cd earthdata-mcp

# Install dependencies
uv sync

# Install dev dependencies (for testing)
uv sync --extra dev

Starting the Local Server

We recommend running the server in HTTP mode for local development and testing:

uv run server.py http

The server will start and be available at http://127.0.0.1:5001/mcp/v1.

Development & Testing

See docs/developers/ for developer guides:

• Adding a New Tool
• Adding an Environment Variable
• Troubleshooting Deployments (Debugging 503 errors, AWS ECS crash loops, and local startup issues)

Running Tests

# Run all tests
uv run pytest

# Run with verbose output
uv run pytest -v

# Run specific test file
uv run pytest tests/test_server.py

Regression Testing

The regression suite runs a golden dataset through a live Bedrock agentic loop and scores results in Langfuse. It requires a running MCP server, AWS credentials, and Langfuse env vars (see .env.example).

Commands

# Push the golden dataset to Langfuse (run once, or after editing mcp_regression_dataset.json)
uv run python evals/regression_test.py publish

# Run experiments against the live dataset and print scores
uv run python evals/regression_test.py run

# Publish then run in one shot
uv run python evals/regression_test.py publish-and-run

Key files

CI gate: set RUN_MCP_REGRESSION=1 to include the regression experiment in pytest.

RUN_MCP_REGRESSION=1 uv run pytest tests/regression/test_regression.py -v

MCP Inspector (Interactive Testing)

1. Start the server:

   复制

   uv run server.py http
   

2. Launch inspector:

   复制

   npx @modelcontextprotocol/inspector
   

3. Connect at http://localhost:6274:

   • Transport Type: Streamable HTTP
   • URL: http://localhost:5001/mcp/v1

Deployment

The application deploys to AWS via Bamboo CI/CD.

See the Developer Guides for detailed architectural breakdowns, deployment procedures, infrastructure definitions (terraform/), and troubleshooting steps for server startup issues.

Publishing to Model Context Protocol Registry

Pre-requisites:

1. Download mcp-publisher from homebrew or make the source on https://github.com/modelcontextprotocol/registry

2. Publishing is handled via GitHub Actions using OIDC authentication — see .github/workflows/publish-mcp.yml. Manual login via mcp-publisher login github currently fails for io.github.nasa/* due to a known registry bug (modelcontextprotocol/registry#1537) where org membership isn't picked up correctly. Use the workflow instead of logging in locally.

Publishing Steps:

1. Since we are supporting multiple MCP servers, first bump the version in server.json — each release is immutable, so increment to a new version number, e.g. "version": "2.0.0", and add a new remote, e.g.

   {
     "type": "streamable-http",
     "url": "https://cmr.earthdata.nasa.gov/mcp/v2"
   }

Note: you cannot update the registry entry without bumping the version number, as the registry is immutable.

2. Commit and get your changes for server.json merged into to main after the PR.

3. Trigger the Publish to MCP Registry workflow from the Actions tab (or gh workflow run publish-mcp.yml). This authenticates via GitHub OIDC and runs mcp-publisher publish automatically — by default this publishes to the MCP registry at https://registry.modelcontextprotocol.io/servers and is publicly available to all MCP clients.