MCP 服务器

source-coop-mcp

Enables AI agents to discover and access 800TB+ of public geospatial data from Source Cooperative, with tools for listing organizations, products, files, and fuzzy search.

README

Source Cooperative MCP Server

Discover and access 800TB+ of geospatial data through AI agents.

An MCP (Model Context Protocol) server for Source Cooperative - a collaborative repository with datasets from Maxar, Harvard, ESA, USGS, and 90+ organizations.

🏗️ Architecture Overview

graph TB
    subgraph "AI Clients"
        A1[Claude Desktop]
        A2[Claude Code]
        A3[Cursor]
        A4[Cline]
        A5[Zed]
        A6[Continue.dev]
    end

    subgraph "MCP Server"
        MCP[Source Cooperative MCP<br/>FastMCP + obstore]
    end

    subgraph "6 Available Tools"
        T1[list_accounts<br/>94+ orgs]
        T2[list_products<br/>hybrid S3+API]
        T3[get_product_details<br/>+ README]
        T4[list_product_files<br/>tree mode]
        T5[get_file_metadata<br/>no download]
        T6[search<br/>hybrid fuzzy]
    end

    subgraph "Data Sources"
        S1[HTTP API<br/>source.coop/api]
        S2[S3 Direct<br/>opendata.source.coop]
    end

    A1 -->|JSON-RPC| MCP
    A2 -->|JSON-RPC| MCP
    A3 -->|JSON-RPC| MCP
    A4 -->|JSON-RPC| MCP
    A5 -->|JSON-RPC| MCP
    A6 -->|JSON-RPC| MCP

    MCP --> T1
    MCP --> T2
    MCP --> T3
    MCP --> T4
    MCP --> T5
    MCP --> T6

    T1 --> S2
    T2 --> S1
    T2 --> S2
    T3 --> S1
    T3 --> S2
    T4 --> S2
    T5 --> S2
    T6 --> S1

    style MCP fill:#4CAF50,stroke:#2E7D32,stroke-width:3px,color:#fff
    style S1 fill:#2196F3,stroke:#1976D2,stroke-width:2px,color:#fff
    style S2 fill:#2196F3,stroke:#1976D2,stroke-width:2px,color:#fff

Key Features:

✅ Token Optimized - 72% reduction for large datasets
✅ Smart Partitions - Auto-detects Hive-style patterns
✅ Fuzzy Search - Handles typos and partial matches
✅ No Auth - All 800TB+ is public

🚀 Quick Start

Install

uvx source-coop-mcp

Configure Your AI Client

Claude Desktop / Claude Code / Cursor / Cline

Add to config file:

Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
Claude Code: VS Code settings.json
Cursor: Cursor settings
Cline: Cline MCP settings

{
  "mcpServers": {
    "source-coop": {
      "command": "uvx",
      "args": ["source-coop-mcp"]
    }
  }
}

Zed

Add to Zed settings:

{
  "context_servers": {
    "source-coop": {
      "command": "uvx",
      "args": ["source-coop-mcp"]
    }
  }
}

Continue.dev

Add to Continue config (~/.continue/config.json):

{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "uvx",
          "args": ["source-coop-mcp"]
        }
      }
    ]
  }
}

Restart your AI client and start exploring!

🛠️ Available Tools

Tool	Purpose	Performance
`list_accounts()`	Find all 94+ organizations	~850ms
`list_products()`	Hybrid: S3 mode (default) for ALL datasets + file counts	~240ms
`list_products(include_unpublished=False)`	API mode for published datasets with rich metadata	~500ms
`get_product_details()`	Get metadata + README automatically	~650ms
`list_product_files()`	List files with S3/HTTP paths	~240ms
`list_product_files(show_tree=True)`	Tree view (72% token savings)	~980ms
`get_file_metadata()`	Get file info without downloading	~230ms
`search(query)`	Hybrid: Search accounts + products (published + unpublished), top 5 results	~5-10s

💡 What You Can Do

Discover Data

"List all organizations in Source Cooperative"
→ Returns 94+ organizations: maxar, planet, harvard, etc.

"Find all datasets for harvard-lil"
→ Discovers published + unpublished products

"Search for climate datasets"
→ Smart fuzzy search handles typos and partial matches

Access Files

"List files in harvard-lil/gov-data"
→ Returns S3 paths and HTTP URLs ready for analysis

"Show me the file tree with partition detection"
→ Smart visualization: year={2020,2021,...+5 more}/ [partitioned]

"Get file metadata without downloading"
→ Size, last modified, ETag

Smart Search

"Search for climte" (typo)
→ Finds "climate" datasets (fuzzy matching)

"Search for geo" (partial)
→ Finds "geospatial", "geocoding", etc.

⚡ Features

Feature	Description
Complete Discovery	Finds unpublished products the official API doesn't show
No Authentication	All 800TB+ data is public
Fast Performance	Rust-backed S3 client (9x faster than boto3)
Token Optimized	Tree mode: 72% token reduction for large datasets
Smart Partitions	Auto-detects patterns: `year={2020,2021,...}`
Fuzzy Search	Handles typos and partial matches
README Integration	Documentation automatically included
800TB+ Data	94+ organizations, geospatial datasets

📋 Example Workflow

1. "List all organizations"
   → Get 94+ account names

2. "Show me all datasets from maxar"
   → Discover published + unpublished products

3. "Search for climate data"
   → Smart fuzzy search finds relevant datasets

4. "Get details for harvard-lil/gov-data"
   → Full metadata + README content

5. "List files in this dataset with tree view"
   → Token-optimized tree with partition detection

🎯 Why This Server?

Problem

Source Cooperative has 800TB+ of valuable data, but:

Official API only shows published products
No auto-discovery of organizations
Requires knowing what you're looking for

Solution

This MCP server provides:

✅ Complete auto-discovery (published + unpublished)
✅ Smart search with fuzzy matching
✅ Direct S3 access for all files
✅ Token-optimized outputs (72% reduction)
✅ Smart partition detection (10-88% additional savings)
✅ README documentation included automatically
✅ No authentication required

📊 Performance

All operations complete in under 1 second:

list_accounts():                          ~850ms  (94+ organizations)
list_products():                          ~240ms  (S3 mode - ALL datasets + file counts)
list_products(include_unpublished=False): ~500ms  (API mode - published with metadata)
list_product_files():                     ~240ms  (simple list)
list_product_files(tree=True):            ~980ms  (72% token savings)
get_file_metadata():                      ~230ms  (HEAD only)
search(query):                            ~5-10s  (hybrid search - 1 recursive S3 scan, top 5 enriched)