source-coop-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@source-coop-mcp search for climate data"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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.
šļø 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:#fffKey Features:
ā Token Optimized - 72% reduction for large datasets
ā Smart Partitions - Auto-detects Hive-style patterns
ā Fuzzy Search - Handles typos and partial matches
ā No Auth - All 800TB+ is public
Related MCP server: SkyFi MCP Server
š Quick Start
Install
uvx source-coop-mcpConfigure 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.jsonCursor: 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!
š ļø Available Tools
Tool | Purpose | Performance |
| Find all 94+ organizations | ~850ms |
| Hybrid: S3 mode (default) for ALL datasets + file counts | ~240ms |
| API mode for published datasets with rich metadata | ~500ms |
| Get metadata + README automatically | ~650ms |
| List files with S3/HTTP paths | ~240ms |
| Tree view (72% token savings) | ~980ms |
| Get file info without downloading | ~230ms |
| Hybrid: Search accounts + products (published + unpublished), top 5 results | ~5-10s |
š” What You Can Do
Discover Data
"List all organizations in Source Cooperative"
ā Returns 94+ organizations: maxar, planet, harvard, etc.
"Find all datasets for harvard-lil"
ā Discovers published + unpublished products
"Search for climate datasets"
ā Smart fuzzy search handles typos and partial matchesAccess Files
"List files in harvard-lil/gov-data"
ā Returns S3 paths and HTTP URLs ready for analysis
"Show me the file tree with partition detection"
ā Smart visualization: year={2020,2021,...+5 more}/ [partitioned]
"Get file metadata without downloading"
ā Size, last modified, ETagSmart Search
"Search for climte" (typo)
ā Finds "climate" datasets (fuzzy matching)
"Search for geo" (partial)
ā Finds "geospatial", "geocoding", etc.ā” 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: |
Fuzzy Search | Handles typos and partial matches |
README Integration | Documentation automatically included |
800TB+ Data | 94+ organizations, geospatial datasets |
š Example Workflow
1. "List all organizations"
ā Get 94+ account names
2. "Show me all datasets from maxar"
ā Discover published + unpublished products
3. "Search for climate data"
ā Smart fuzzy search finds relevant datasets
4. "Get details for harvard-lil/gov-data"
ā Full metadata + README content
5. "List files in this dataset with tree view"
ā Token-optimized tree with partition detectionšÆ 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:
ā Complete auto-discovery (published + unpublished)
ā Smart search with fuzzy matching
ā Direct S3 access for all files
ā Token-optimized outputs (72% reduction)
ā Smart partition detection (10-88% additional savings)
ā README documentation included automatically
ā No authentication required
š 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)Token Optimization Impact
Dataset Size | Without Tree | With Tree | Saved |
10 files | 1,500 tokens | 415 tokens | 72.3% |
100 files | 15,000 tokens | 4,150 tokens | 72.3% |
1,000 files | 150,000 tokens | 41,500 tokens | 72.3% |
With partition detection (1,000 partitions): 88% total savings!
š§ Requirements
Python: 3.11 or higher
Package Manager:
uv(installed automatically byuvx)Operating Systems: macOS, Linux, Windows
š¤ Development
See DEVELOPMENT.md for:
Architecture details
Testing instructions
Contributing guidelines
Performance benchmarks
Token optimization details
š Support
Issues: GitHub Issues
š License
MIT License - see LICENSE for details.
Available Tools
6 toolsget_file_metadataA
Get metadata for a specific file without downloading it. Uses obstore's head operation for efficient metadata retrieval.
Args: path: S3 URI (s3://...) or relative path (account_id/product_id/file)
Returns: File metadata: size, content-type, last-modified, etag, URLs
Example: >>> await get_file_metadata("harvard-lil/gov-data/README.md") { "key": "harvard-lil/gov-data/README.md", "content_type": "binary/octet-stream", "content_length": 5344, "last_modified": "2025-02-06T16:29:24+00:00", ... }
Inputs: path (required). Returns: object
get_product_detailsA
Get comprehensive metadata for a specific product. Always includes README content if found in the product root directory.
Args: account_id: Account ID (e.g., "harvard-lil") product_id: Product ID (e.g., "gov-data")
Returns: Full product metadata including account info, storage config, roles, tags Always includes 'readme' field with content and metadata (if README exists)
Example: >>> await get_product_details("harvard-lil", "gov-data") { "title": "Archive of data.gov", "description": "...", "account": {"name": "Harvard Library Innovation Lab", ...}, "readme": { "found": true, "content": "# Archive of data.gov...", "size": 5344, "path": "harvard-lil/gov-data/README.md" }, ... }
Inputs: account_id (required), product_id (required). Returns: object
list_accountsA
Discover all organizations/accounts in Source Cooperative.
Returns: List of account IDs (e.g., ['clarkcga', 'harvard-lil', 'youssef-harby'])
Example: >>> await list_accounts() ['addresscloud', 'clarkcga', 'harvard-lil', ...]
Inputs: no named parameters. Returns: result
list_product_filesA
List all files in a product with full S3 paths ready for analysis. Optionally show a hierarchical tree visualization (optimized for LLM tokens).
Args: account_id: Account ID product_id: Product ID prefix: Optional prefix to filter files (subdirectory path) max_files: Maximum files to return (default 1000) show_tree: If True, return tree visualization only (more token-efficient, default True)
Returns: Dict with either files list OR tree visualization (not both to save tokens)
Example (List mode - detailed metadata): >>> result = await list_product_files("harvard-lil", "gov-data", "metadata/") >>> print(result["files"][0]) { "key": "harvard-lil/gov-data/metadata/metadata.jsonl.zip", "s3_uri": "s3://us-west-2.opendata.source.coop/harvard-lil/gov-data/metadata/metadata.jsonl.zip", "http_url": "https://data.source.coop/harvard-lil/gov-data/metadata/metadata.jsonl.zip", "size": 1012127330, "last_modified": "2025-02-06T16:20:22+00:00" }
Example (Tree mode - token optimized): >>> result = await list_product_files("harvard-lil", "gov-data", show_tree=True) >>> print(result["tree"]) s3://us-west-2.opendata.source.coop/harvard-lil/gov-data/ āāā README.md (5.2 KB) ā s3://...README.md āāā metadata/ ā āāā metadata.jsonl.zip (965.4 MB) ā s3://...metadata.jsonl.zip āāā data/ āāā datasets.parquet (128.5 MB) ā s3://...datasets.parquet
Example (Partitioned data - smart summarization): >>> result = await list_product_files("account", "product", show_tree=True) >>> print(result["tree"]) s3://us-west-2.opendata.source.coop/account/product/ āāā year={1995,1996,...,2007 (13 total)}/ [partitioned] ā āāā format={ixi,pxp}/ [partitioned] ā āāā matrix={F_impacts,F_satellite,Y,Z}/ [partitioned] ā āāā data.parquet (5.1 MB)
Note: Shows first,second,...,last (total) for >10 values; lists all for ā¤10
Tree mode saves ~70% tokens + smart partition detection saves 96%+ moreInputs: account_id (required), product_id (required), prefix, max_files, +1 more. Returns: object
list_productsA
List products (datasets) in Source Cooperative with hybrid S3 + API approach.
DEFAULT: Uses S3 direct scan (fast, includes ALL products with file counts). Set include_unpublished=False for published-only with rich metadata from API.
Args: account_id: Filter by specific account. REQUIRED for S3 mode (default). If None with include_unpublished=False, lists published from all accounts. featured_only: Only return featured/curated products (API mode only). include_unpublished: If True (default), scan S3 for ALL products including unpublished. If False, use API for published products with rich metadata. include_file_count: Count files in each product (default True, only in S3 mode).
Returns: S3 mode (default): Basic info (product_id, s3_prefix, file_count) - fast! API mode: Rich metadata (product_id, title, description, dates) - slower
Performance: - S3 mode (default): ~240ms, includes unpublished products + file counts - API mode (include_unpublished=False): ~500ms, rich metadata, published only
Examples: >>> # ALL products with file counts (DEFAULT - fast!) >>> await list_products(account_id="youssef-harby") [ {"product_id": "exiobase-3", "source": "s3", "file_count": 1000, ...}, {"product_id": "egms-copernicus", "source": "s3", "file_count": 53, ...}, ... ]
>>> # Published products with rich metadata (API mode)
>>> await list_products(account_id="youssef-harby", include_unpublished=False)
[{"product_id": "egms-copernicus", "title": "...", "description": "...", ...}]
>>> # Fast mode without file counts
>>> await list_products(account_id="youssef-harby", include_file_count=False)
[{"product_id": "exiobase-3", "source": "s3", ...}]
>>> # Featured products only (requires API mode)
>>> await list_products(featured_only=True, include_unpublished=False)
[{"product_id": "gov-data", "featured": 1, ...}]Inputs: account_id, featured_only, include_unpublished, include_file_count. Returns: result
searchA
Search for products across ALL accounts with smart fuzzy matching. Handles typos, partial matches, and incomplete words using 60% similarity threshold.
Hybrid Search - Automatically searches across:
All 94+ organizations
ALL products (published + unpublished)
All fields: title, description, product_id
Published products: Full metadata (title, description, product_id) Unpublished products: product_id only (no title/description available)
Args: query: Search keyword (supports typos and partial matches)
Returns: Top 5 matching accounts or products (sorted by relevance score)
Performance: ~5-8s (parallel 2-level S3 scan + top 5 API enrichment)
Performance breakdown:
- S3 parallel listing: ~2.4s (94 accounts + 354 products)
- Fuzzy matching: <1s (in-memory processing)
- API enrichment: ~2-5s (only top 5 results)
**11x faster** than sequential approach (was ~27s)
**Uses 2-level delimiter listing** (not full recursive scan)Examples: >>> # Exact match >>> results = await search("climate")
>>> # Fuzzy match (handles typos)
>>> results = await search("climte") # Finds "climate"
>>> results = await search("exiopase") # Finds "exiobase-3" (includes unpublished!)
>>> # Partial match
>>> results = await search("geo") # Finds "geospatial", "geocoding", etc.
>>> # Result formats
>>> print(results[0]) # Account match
{
"type": "account",
"account_id": "harvard-lil",
"match_string": "harvard-lil",
"search_score": 9.5,
"similarity": 0.95,
"matched_fields": ["account_id"]
}
>>> print(results[1]) # Product match
{
"type": "product",
"account_id": "youssef-harby",
"product_id": "exiobase-3",
"match_string": "youssef-harby/exiobase-3",
"title": "", # Empty for unpublished products
"description": "", # Empty for unpublished products
"search_score": 8.2,
"similarity": 0.82,
"matched_fields": ["product_id"]
}Inputs: query (required). Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelogā¦
TDQS
Scored across 6 tools
Each tool has a clear, distinct purpose: metadata retrieval for files vs. products, listing accounts, products, or files, and a separate search function. No overlap in functionality.
All tool names follow a consistent verb_noun pattern (get_file_metadata, list_accounts, search, etc.) using lowercase with underscores. Highly predictable and uniform.
Six tools is ideal for a data catalog server, covering discovery, browsing, and search without being overwhelming or too sparse.
The tool surface covers all essential operations for exploring a data cooperative: listing accounts, products, files, retrieving detailed metadata, and cross-account search. No obvious gaps.
Maintenance
Related MCP Connectors
AI Agent Source Registry. 288K+ curated sources for agentic search and discovery.
Geodata for any place, inside your agent: map data, terrain, routes, satellite images, GIS files.
Connect AI agents to primary sources - markets, macro, news, earth, network, entities and research.
SEC filings, Wikipedia, arXiv, US Census and more for AI agents, with every answer source-linked.
Related MCP Servers
- AlicenseBqualityCmaintenanceEnables AI assistants to search and access geospatial datasets through STAC (SpatioTemporal Asset Catalog) APIs. Supports querying satellite imagery, weather data, and other geospatial assets with spatial, temporal, and attribute filters.1120MIT
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to search, compare pricing, and order satellite imagery from 150+ satellites across 12+ providers via natural language.MIT
- AlicenseAqualityBmaintenanceEnables AI agents to query Earth observation data, satellite imagery, active fires, and weather via natural language, returning interactive maps and briefs.12MIT
- FlicenseNot gradedqualityDmaintenanceTurns the AWS Registry of Open Data and public STAC catalogs into agent-callable tools for dataset search, STAC queries, and NDVI/NDBI/land-cover analysis on Sentinel-2 imagery.-