stac-mcp 1.2.3

An MCP Server for STAC requests

STAC MCP Server

An MCP (Model Context Protocol) Server that provides access to STAC (SpatioTemporal Asset Catalog) APIs for geospatial data discovery and access. Supports dual output modes (text and structured json) for all tools.

The coverage badge is updated automatically on pushes to main by the CI workflow.

Overview

This MCP server enables AI assistants and applications to interact with STAC catalogs to:

• Search and browse STAC collections
• Find geospatial datasets (satellite imagery, weather data, etc.)
• Access metadata and asset information
• Perform spatial and temporal queries

Features

Available Tools

All tools accept an optional output_format parameter ("text" default, or "json"). JSON mode returns a single MCP TextContent whose text field is a compact JSON envelope: { "mode": "json", "data": { ... } } (or { "mode": "text_fallback", "content": ["..."] } if a handler lacks a JSON branch). This preserves backward compatibility while enabling structured consumption (see ADR 0006 and ASR 1003).

• get_root: Fetch root document (id/title/description/links/conformance subset)
• get_conformance: List all conformance classes; optionally verify specific URIs
• get_queryables: Retrieve queryable fields (global or per collection) when supported
• get_aggregations: Execute a search requesting aggregations (count/stats) if supported
• search_collections: List and search available STAC collections
• get_collection: Get detailed information about a specific collection
• search_items: Search for STAC items with spatial, temporal, and attribute filters
• get_item: Get detailed information about a specific STAC item
• estimate_data_size: Estimate data size for STAC items using lazy loading (XArray + odc.stac)

PySTAC-based CRUDL Tools

The server also provides a comprehensive set of PySTAC-based tools for managing STAC resources with full Create, Read, Update, Delete, and List (CRUDL) operations. These tools work with both local filesystems and remote STAC servers:

Catalog Management:

• pystac_create_catalog: Create a new STAC Catalog (local or remote)
• pystac_read_catalog: Read an existing STAC Catalog
• pystac_update_catalog: Update a STAC Catalog
• pystac_delete_catalog: Delete a STAC Catalog
• pystac_list_catalogs: List STAC Catalogs in a directory or endpoint

Collection Management:

• pystac_create_collection: Create a new STAC Collection (local or remote)
• pystac_read_collection: Read an existing STAC Collection
• pystac_update_collection: Update a STAC Collection
• pystac_delete_collection: Delete a STAC Collection
• pystac_list_collections: List STAC Collections in a directory or endpoint

Item Management:

• pystac_create_item: Create a new STAC Item (local or remote)
• pystac_read_item: Read an existing STAC Item
• pystac_update_item: Update a STAC Item
• pystac_delete_item: Delete a STAC Item
• pystac_list_items: List STAC Items in a directory or endpoint

Key Features:

• Local Operations: Manage STAC resources on the filesystem
• Remote Operations: Interact with remote STAC servers via HTTP/HTTPS
• API Key Authentication: Set STAC_API_KEY environment variable for authenticated requests
• Complementary: Works alongside existing transaction tools without replacing them

For detailed documentation and examples, see PySTAC CRUDL Tools Documentation.

Capability Discovery & Aggregations

The new capability tools (ADR 0004) allow adaptive client behavior:

• Graceful fallbacks: Missing /conformance, /queryables, or aggregation support returns structured JSON with supported:false instead of hard errors.
• get_conformance falls back to the root document's conformsTo array when the dedicated endpoint is absent.
• get_queryables returns an empty set with a message if the endpoint is not implemented by the catalog.
• get_aggregations constructs a STAC Search request with an aggregations object; if unsupported (HTTP 400/404), it returns a descriptive message while preserving original search parameters.

Data Size Estimation

The estimate_data_size tool provides accurate size estimates for geospatial datasets without downloading the actual data:

• Lazy Loading: Uses odc.stac to load STAC items into xarray datasets without downloading
• AOI Clipping: Automatically clips to the smallest area when both bbox and AOI GeoJSON are provided
• Fallback Estimation: Provides size estimates even when odc.stac fails
• Detailed Metadata: Returns information about data variables, spatial dimensions, and individual assets
• Batch Support: Retains structured metadata for efficient batch processing

Usage

MCP Protocol / Server Configuration

The server implements the Model Context Protocol (MCP) for standardized communication.

{
  "stac": {
    "command": "uvx",
    "args": [
      "--from",
      "git+https://github.com/wayfinder-foundry/stac-mcp",
      "stac-mcp"
    ],
    "transport": "stdio",
  }
}

Command Line

Native Installation

pip install stac-mcp

stac-mcp

Each invocation starts an MCP stdio server; it waits for protocol messages (see examples/example_usage.py).

Repository Usage

pip install -e .

Container Usage

Local

docker build -t stac-mcp .

docker run --rm -i stac-mcp

Published Image

# With Docker
docker run --rm -i ghcr.io/bnjam/stac-mcp:latest

# With Podman
podman run --rm -i ghcr.io/bnjam/stac-mcp:latest

Examples

Example: JSON Output Mode

Below is an illustrative (client-side) pseudo-call showing output_format usage through an MCP client message:

{
  "method": "tools/call",
  "params": {
    "name": "search_items",
    "arguments": {
      "collections": ["landsat-c2l2-sr"],
      "bbox": [-122.5, 37.7, -122.3, 37.8],
      "datetime": "2023-01-01/2023-01-31",
      "limit": 5,
      "output_format": "json"
    }
  }
}

The server responds with a single TextContent whose text is a JSON string like:

{"mode":"json","data":{"type":"item_list","count":5,"items":[{"id":"..."}]}}

This wrapping keeps the MCP content type stable while enabling machine-readable chaining.

Development

Setup

GitHub Codespaces (Recommended)

The fastest way to get started is with GitHub Codespaces, which provides a fully configured development environment in your browser:

1. Click the green "Code" button on the GitHub repository
2. Select the "Codespaces" tab
3. Click "Create codespace on main"

The devcontainer will automatically:

• Set up Python 3.12 with all dependencies
• Install GDAL/PROJ system libraries
• Configure VS Code with recommended extensions
• Install the project in development mode

See .devcontainer/README.md for more details.

Local Development

git clone https://github.com/BnJam/stac-mcp.git
cd stac-mcp
pip install -e ".[dev]"

For local development with containers, you can use VS Code's Remote Containers extension with the provided .devcontainer configuration.

Testing

pytest -v
python examples/example_usage.py  # MCP stdio smoke test

Test Coverage

The project uses coverage.py (already a dependency was added) for measuring statement and branch coverage.

Quick run (terminal):

coverage run -m pytest -q
coverage report -m

Example output (illustrative):

Name                                Stmts   Miss Branch BrMiss  Cover
---------------------------------------------------------------------
stac_mcp/observability.py             185      4     42      3    96%
stac_mcp/tools/execution.py            68      2     18      1    94%
... (others) ...
---------------------------------------------------------------------
TOTAL                                 620     20    140      9    96%

Generate an HTML report (optional):

coverage html
open htmlcov/index.html  # macOS

Configuration: .coveragerc enforces branch = True and omits tests/* and scripts/version.py. Update omit patterns only when necessary to keep metrics honest.

Recommended workflow before opening a PR:

1. ruff format stac_mcp/ tests/ examples/
2. ruff check stac_mcp/ tests/ examples/ --fix
3. coverage run -m pytest -q
4. coverage report -m (ensure no unexpected drops)

SSL / TLS Troubleshooting

If you encounter an SSL certificate verification error (e.g., SSLCertVerificationError or a message about a self-signed certificate in certificate chain) when the server accesses a STAC endpoint:

1. Confirm the endpoint is reachable with a standard tool (e.g., curl https://.../stac/v1/conformance).
2. Ensure your system trust store is up to date (on macOS, some Python installs provide an Install Certificates.command).
3. Behind a corporate proxy / MITM device: export a custom CA bundle.

The client now supports two environment variables (see ADR notes / security guidance):

Variable	Purpose	Security Impact
STAC_MCP_CA_BUNDLE	Path to a PEM file with additional / custom root CAs. If present and readable it will be used to build the SSL context.	Low (extends trust roots intentionally).
STAC_MCP_UNSAFE_DISABLE_SSL	If set to 1, disables certificate verification entirely (hostname + chain). For diagnostics only.	High (vulnerable to MITM). Never use in production.

Example (custom CA):

export STAC_MCP_CA_BUNDLE=/etc/ssl/certs/internal-proxy.pem
stac-mcp

Temporary diagnostic bypass (NOT recommended):

export STAC_MCP_UNSAFE_DISABLE_SSL=1
stac-mcp

When an SSL failure occurs you will receive a structured SSLVerificationError message with remediation guidance instead of a low-level urllib.error.URLError.

Container vs Local/Virtual Environment (Why get_conformance May Differ)

The published Docker/Podman images generally succeed with get_conformance against public STAC APIs even when a locally installed Python environment fails. Reasons:

• The container base image (python:3.12-slim) ships with a current CA trust store.
• Some local macOS / Homebrew / pyenv environments have an out-of-date or un-initialized certificate bundle until you run the platform's certificate installation script.
• Corporate proxies can inject custom CAs that exist in system Keychain but are not automatically propagated to the Python cert store.

Typical symptom: Local invocation of the get_conformance tool returns a structured SSLVerificationError, while running the same command via the container (e.g. docker run --rm -i ghcr.io/bnjam/stac-mcp:latest) succeeds.

Mitigations (ordered):

1. Update local certificates (macOS framework Python: run the Install Certificates.command script found in the Python application folder).
2. Export a custom CA bundle path: export STAC_MCP_CA_BUNDLE=/path/to/ca.pem.
3. (Last resort, diagnostics only) Temporarily disable verification with STAC_MCP_UNSAFE_DISABLE_SSL=1 and immediately revert once the root cause is identified.

If the container also fails, the remote endpoint may genuinely present an invalid or mismatched certificate—collect the structured error details (they include hostname and failing reason) and investigate network or proxy layers.

Planned future enhancements (pending ADRs): add retry/federation logic and corresponding tests; coverage thresholds may be introduced once feature set stabilizes.

Linting

ruff format stac_mcp/ tests/ examples/
ruff check stac_mcp/ tests/ examples/

Version Management

The project uses semantic versioning (SemVer) with automated version management based on PR labels or branch naming, implemented in .github/workflows/container.yml.

Automatic Versioning

When PRs are merged to main, the workflow determines the version increment using either PR labels or branch prefixes:

PR Labels (Recommended for Automated Tools)

Labels take priority over branch prefixes. Add one of these labels to your PR:

• bump:patch or bump:hotfix → patch increment (0.1.0 → 0.1.1) for bug fixes
• bump:minor or bump:feature → minor increment (0.1.0 → 0.2.0) for new features
• bump:major or bump:release → major increment (0.1.0 → 1.0.0) for breaking changes

Branch Prefixes (For Human Contributors)

If no version bump label is present, the workflow falls back to branch prefix detection:

• hotfix/, fix/, copilot/fix-, or copilot/hotfix/ branches → patch increment (0.1.0 → 0.1.1) for bug fixes
• feature/ or copilot/feature/ branches → minor increment (0.1.0 → 0.2.0) for new features
• release/ or copilot/release/ branches → major increment (0.1.0 → 1.0.0) for breaking changes

See CONTRIBUTING.md for detailed guidelines on version bumping.

Manual Version Management

You can also manually manage versions using the version script (should normally not be needed unless doing a coordinated release):

# Show current version
python scripts/version.py current

# Increment version based on change type
python scripts/version.py patch    # Bug fixes (0.1.0 -> 0.1.1)
python scripts/version.py minor    # New features (0.1.0 -> 0.2.0)  
python scripts/version.py major    # Breaking changes (0.1.0 -> 1.0.0)

# Set specific version
python scripts/version.py set 1.2.3

The version system maintains consistency across:

• pyproject.toml (project version)
• stac_mcp/__init__.py (version)
• stac_mcp/fast_server.py (server_version in MCP initialization)

Container Development

To develop with containers:

# Build development image
docker build -f Containerfile -t stac-mcp:dev .

# Test the container
docker run --rm -i stac-mcp:dev

# Using docker-compose for development
docker-compose up --build

# For debugging, use an interactive shell (requires modifying Containerfile)
# docker run --rm -it --entrypoint=/bin/sh stac-mcp:dev

Current Containerfile (single-stage) notes:

• Based on python:3.12-slim for broad wheel compatibility (rasterio, shapely, etc.)
• Installs GDAL/PROJ system libraries needed by rasterio/odc-stac
• Installs the package with pip install .
• Entrypoint: python -m stac_mcp.server (stdio MCP transport)
• Multi-stage/distroless hardening can be reintroduced later (tracked by potential future ADR)

Documentation

FastMCP Guidelines and Architecture

STAC MCP includes comprehensive documentation for FastMCP patterns and agentic geospatial reasoning:

• FastMCP Documentation: Complete guide to MCP decorators, resources, tools, and prompts for STAC workflows
  • DECORATORS.md: Choosing the right decorator for STAC operations
  • GUIDELINES.md: FastMCP architecture and usage patterns
  • PROMPTS.md: Agentic STAC search reasoning and methodology
  • RESOURCES.md: STAC catalog discovery and metadata patterns
  • CONTEXT.md: Context usage for logging and progress tracking

These documents provide guidance for:

• AI agents reasoning about STAC catalog searches
• Developers implementing STAC MCP features
• Understanding the planned FastMCP integration (issues #69, #78)

Additional Documentation

• PySTAC CRUDL Tools Documentation: Detailed guide to PySTAC-based CRUD operations
• Test Coverage Strategy: Testing approach and coverage goals

STAC Resources

• STAC Specification
• pystac-client Documentation
• Microsoft Planetary Computer
• AWS Earth Search

License

Apache 2.0 - see LICENSE file for details.

Architecture Overview

The project maintains Architecture Decision Records (ADRs) and Architecture Significant Requirements (ASRs) under architecture/. Core recent decisions:

• Observability & Telemetry (ADR 0012): structured logging (stderr only), metrics counters, correlation IDs, future-ready tracing hooks.
• Multi-Catalog Federation (ADR 0013): optional parallel search across multiple STAC endpoints with deterministic merging and provenance.
• Pluggable Tool Extension Model (ADR 0014): entry point / directory-based plugin registration with collision protection.
• Response Meta Stability (ADR 0015): introduces meta object with stable vs experimental field tiering.
• Security & Credential Isolation (ADR 0016): alias-scoped credentials, redaction and least-privilege injection.

Notable earlier foundations:

• Output format & JSON envelope (ADR 0006) and JSON stability (ASR 1003)
• Capability & aggregation support (ADR 0004)
• Data size estimation tool (ADR 0009) with nodata efficiency requirement (ASR 1006)
• Caching layer (ADR 0011)
• Offline deterministic validation (ASR 1001)
• Graceful network error handling (ASR 1004)
• Performance bounds for search (ASR 1005)
• Reliability & Retry Policy (ASR 1008)

See individual ADR/ASR markdown files for full context, rationale, and evolution notes.

Service Level Objectives (SLO) & Requirements Summary

The following summarizes measurable targets defined in ASRs (and related ADR enforcement points). These are engineering goals; enforcement is via tests, benchmarks, and observability counters.

Area	Reference	Objective
Offline Dev & Tests	ASR 1001	Install <=120s, tests <=30s, example script ~0.6s, no live network
JSON Output Stability	ASR 1003	Backwards-compatible JSON schemas within major version; golden tests guard
Network Error Handling	ASR 1004	All network faults mapped to structured errors; server never crashes
Search Performance Bounds	ASR 1005	Conservative default limit (10); pagination controls; no unbounded iteration
NoData & Memory Efficiency	ASR 1006	Optional adjusted size reporting with adjust_for_nodata; always provide raw & adjusted bytes
Reliability & Retries	ASR 1008	>=95% success under 20% transient fault injection; p95 retry overhead <= +35%; max invocation 15s; ≤2 retries (3 attempts total)
Meta Stability	ADR 0015	Stable vs _exp_ field tiering; no breaking removal of stable fields within major version
Observability	ADR 0012	Structured JSON logs (opt-in), correlation IDs per request, metrics counters (latency, errors, cache, retries)
Federation	ADR 0013	Partial catalog failures produce warnings not total failure when at least one succeeds
Plugin Safety	ADR 0014	Tool name collision prevention; load failures isolated; optional strict mode
Credential Isolation	ADR 0016	Per-alias credential scoping; automatic redaction; plugin access opt-in

Experimental Meta Fields (Subject to Change)

Defined in ADR 0015; current experimental keys returned (when features enabled):

• _exp_federation_warnings: array of partial-failure or truncation notices
• _exp_cache_hit: boolean indicating cache usage
• _exp_retry_attempts: integer number of retry attempts performed

Promotion of experimental fields to stable requires an ADR update and minor version release; consumers should treat _exp_* names as best-effort hints.

Operational Notes

• Logging never uses stdout to avoid MCP protocol interference (ADR 0012).
• Federation item merging adds provenance via a namespaced property (stac_mcp:source_catalog) (ADR 0013).
• Retry logic applies only to idempotent read tools; future write-type tools must opt in explicitly (ASR 1008).
• Nodata adjustment is off by default to preserve raw size semantics (ASR 1006).

Roadmap Candidates (Future ADRs)

• Metrics exposure tool or external exporter integration
• Circuit breaker & adaptive backoff extensions to reliability policy
• Plugin capability introspection tool
• OAuth / token refresh flows for credential layer

For contributions impacting architecture, add or update an ADR/ASR following AGENTS.md guidelines.

Client Configuration (ADR 0007)

The STAC client implementation supports flexible configuration options for varied deployment scenarios (corporate proxies, slow networks, custom authentication).

Per-Call Configuration (Programmatic API)

When using the STACClient class directly (e.g., in custom tools or extensions), you can configure requests at call time:

Timeout Configuration

All STAC API requests support an optional timeout parameter (in seconds):

from stac_mcp.tools.client import STACClient

client = STACClient("https://planetarycomputer.microsoft.com/api/stac/v1")

# Use custom timeout (60 seconds instead of default 30)
result = client._http_json("/conformance", timeout=60)

# Disable timeout (wait indefinitely)
result = client._http_json("/conformance", timeout=0)

Default: 30 seconds if not specified

Use cases:

• Slow networks or high-latency connections: increase timeout
• Large catalog queries: increase timeout to prevent premature failures
• Testing/diagnostics: adjust timeout to isolate performance issues

Headers Configuration

Custom headers can be provided at two levels:

1. Instance-level (applies to all requests from that client):

client = STACClient(
    "https://example.com/stac/v1",
    headers={"X-API-Key": "your-key", "User-Agent": "MyApp/1.0"}
)

2. Per-call (merges with or overrides instance headers):

# Override specific header for this call only
result = client._http_json("/search", headers={"X-API-Key": "different-key"})

Behavior: Per-call headers are merged with instance headers, with per-call values taking precedence for duplicate keys. The Accept: application/json header is always set automatically.

Note: These configuration options are for programmatic use. MCP tool calls use the default client configuration.

HEAD request tuning (estimate_data_size fallback)

The estimate_data_size tool uses metadata-first heuristics but will fall back to HTTP HEAD requests to collect Content-Length when metadata is missing. These HEAD requests are timeboxed and parallelized; tune the behavior with the following environment variables:

• STAC_MCP_HEAD_TIMEOUT_SECONDS (default: 20)
  • Per-request timeout (in seconds) used for HTTP HEAD requests during fallback estimation.
  • Lower this value to fail fast against unresponsive hosts (for example, 5).
• STAC_MCP_HEAD_MAX_WORKERS (default: 4)
  • Number of concurrent HEAD requests used when the estimator needs to probe multiple asset HREFs.
  • Raising this value reduces wall-clock time for estimation at the cost of more concurrent connections.

Agent tuning guidance:

• For short-running agent chains that need quick, best-effort estimates, set STAC_MCP_HEAD_TIMEOUT_SECONDS=5 and STAC_MCP_HEAD_MAX_WORKERS=8.
• For conservative workloads (avoid load on remote catalogs), keep STAC_MCP_HEAD_MAX_WORKERS small (1-4) and use a moderate timeout (10-20s).

You can also force metadata-only estimation from clients by passing force_metadata_only=True to estimate_data_size, which avoids HEAD/zarr inspection entirely.

These knobs let agent orchestrators choose the right balance between responsiveness and thoroughness.

Error Handling

The client provides structured error types with actionable guidance:

Timeout Errors

When requests exceed the timeout threshold, a STACTimeoutError is raised with context:

Request to https://example.com/stac/v1/search timed out after 30s (attempted 3 times).
Consider increasing timeout parameter or checking network connectivity.

Automatic retries: Timeout errors are retried 3 times with exponential backoff before failing.

Connection Errors

Connection failures are mapped to specific, actionable messages via ConnectionFailedError:

• DNS failures: "DNS lookup failed for [url]. Check the catalog URL and network connectivity."
• Connection refused: "Connection refused by [url]. The server may be down or the URL may be incorrect."
• Network unreachable: "Network unreachable for [url]. Check network connectivity and firewall settings."
• Generic errors: Includes the underlying error reason with remediation guidance

Automatic retries: Connection errors are retried 3 times with exponential backoff (0.2s, 0.4s, 0.8s delays).

SSL/TLS Errors

SSL certificate verification failures raise SSLVerificationError with detailed remediation steps. See the SSL / TLS Troubleshooting section for environment variables and configuration options.

Error Logging

Network errors are logged at ERROR level (not EXCEPTION level) to preserve context without excessive stack traces:

ERROR stac_mcp.tools.client: Connection failed after 3 attempts: DNS lookup failed for ...

This follows ADR 0007 guidance: "Log at error level; no prints."

Observability Configuration (ADR 0012)

Environment variables controlling telemetry:

Variable	Default	Description
STAC_MCP_LOG_LEVEL	WARNING	Logging level (DEBUG, INFO, etc.)
STAC_MCP_LOG_FORMAT	text	Set to json for structured single-line JSON logs
STAC_MCP_ENABLE_METRICS	true	Disable (false) to skip counter increments
STAC_MCP_ENABLE_TRACE	false	Enable lightweight span timing debug logs

All logs are emitted to stderr only; stdout is reserved strictly for MCP protocol traffic. JSON logs include fields: timestamp, level, message, plus optional event, tool_name, duration_ms, error_type, correlation_id, cache_hit, catalog_url.