Geospatial AI with epistemic reasoning

GDAL MCP is a Model Context Protocol (MCP) server that provides AI agents with geospatial analysis capabilities while requiring them to justify their methodological choices through a reflection middleware system.

🎉 v1.1.1 Released (2025-10-26) — Vector tool parity + cross-domain reflection validated
🧠 Reflection System — Domain-based epistemic reasoning that transcends data types
⚡ 75% Cache Hit Rate — Methodology reasoning carries across raster ↔ vector operations

• Quick Start - Installation, setup, and MCP configuration
• Tools Reference - Complete tool documentation with examples
• Vision - Long-term roadmap and philosophy
• Changelog - Release history and updates

Most AI tool systems execute operations immediately when requested. GDAL MCP requires the AI to justify methodological decisions before execution, creating a conversation about the "why" rather than just executing the "what."

Traditional AI tool approach:

User: "Reproject this DEM to Web Mercator"
AI: *silently executes operation*
Result: DEM reprojected (but with what resampling? why that CRS? what tradeoffs?)

GDAL MCP with reflection:

User: "Reproject this DEM to Web Mercator for my web map"
AI: *reflection middleware intercepts*
AI: "Before reprojecting to EPSG:3857, let me justify this CRS choice:
     
     Intent: Web Mercator preserves angles (conformal projection),
             essential for web tile rendering and navigation
     
     Alternatives considered:
     - EPSG:4326 (WGS84) → rejected: not conformal, distorts shapes
     - Local UTM → rejected: limited extent, not web-standard
     
     Tradeoffs: Area distortion increases at high latitudes
                (acceptable for web visualization)
     
     Confidence: High - this is the standard for web mapping"

*justification stored, operation proceeds*
Result: DEM reprojected with documented methodology

1. Utilizes Full MCP Protocol Capabilities

The Model Context Protocol isn't just about tool execution—it's about agentic reasoning and interaction. The reflection system leverages MCP's design to enable:

• Pre-execution prompting (tool dependencies)
• Structured reasoning (schema-validated justifications)
• Stateful workflows (justification caching)
• Human-in-the-loop interaction (advisory prompts)

2. Prevents Silent Failures

Geospatial operations can execute successfully while producing methodologically incorrect results:

• Nearest-neighbor resampling on continuous elevation data (creates artifacts)
• Web Mercator for area calculations (40%+ distortion possible)
• Bilinear interpolation on categorical data (creates invalid class values)

The reflection system surfaces these choices for validation.

3. Educational, Not Restrictive

The AI isn't blocked from executing operations—it's required to demonstrate understanding:

• First use: Explains reasoning, teaches methodology
• Cached: Instant execution (knowledge persists)
• Result: 75%+ cache hit rates, minimal friction

4. Creates Audit Trail

Every methodological decision is documented with:

• Intent (what property must be preserved?)
• Alternatives (what else was considered?)
• Rationale (why this choice?)
• Tradeoffs (what are the limitations?)
• Confidence (high/medium/low)

This enables reproducible geospatial science.

User: "I need to reproject this DEM to UTM for accurate slope analysis,
       then reproject this vector layer to the same CRS for overlay"

AI Workflow:
1. Inspects DEM metadata (raster_info)
2. REFLECTION: Justifies UTM Zone 10N choice (accurate distance/area)
3. REFLECTION: Justifies cubic resampling (smooth gradients for derivatives)
4. Reprojects DEM (raster_reproject)
5. Inspects vector metadata (vector_info)
6. CACHE HIT: Reuses UTM justification (cross-domain!)
7. Reprojects vector (vector_reproject) - instant, no re-prompting
8. Both datasets now aligned in UTM Zone 10N

Result: 2 operations, 2 reflections (not 3!)
Cache hit rate: 50% → Saves time, maintains methodology

The Key Innovation: The CRS justification from step 2 is reused in step 6 because the methodology (why UTM Zone 10N?) is domain-based, not tool-based. It doesn't matter if you're working with raster or vector data—the projection choice reasoning is the same.

See Tools Reference for detailed examples of all available tools.

• Pre-execution reasoning for CRS selection, resampling methods
• Structured justifications (intent, alternatives, choice, tradeoffs, confidence)
• Persistent cache with 75% hit rates in multi-operation workflows
• Cross-domain cache sharing - CRS justification works for both raster AND vector

• Raster tools: info, convert, reproject, stats
• Vector tools: info, reproject, convert, clip, buffer, simplify
• See Tools Reference for complete documentation

• Full type safety (mypy strict mode)
• 72 passing tests
• Workspace security (path validation middleware)
• Python-native (Rasterio/PyProj/pyogrio)
• Real-time feedback via FastMCP Context API

• Workspace catalog for autonomous file discovery
• Metadata intelligence for format detection
• Reference knowledge base (CRS, resampling methods, compression options)

# Run directly from PyPI
uvx --from gdal-mcp gdal --transport stdio

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "gdal-mcp": {
      "command": "uvx",
      "args": ["--from", "gdal-mcp", "gdal", "--transport", "stdio"],
      "env": {
        "GDAL_MCP_WORKSPACES": "/path/to/your/geospatial/data"
      }
    }
  }
}

See QUICKSTART.md for:

• Alternative installation methods (Docker, local development)
• Detailed MCP client configuration
• Workspace security setup
• Troubleshooting guide

GDAL MCP provides 12 production-ready tools across three categories:

• raster_info - Inspect metadata (CRS, resolution, bands, nodata)
• raster_convert - Format conversion with compression & overviews (COG support)
• raster_reproject ⚡ - CRS transformation (with reflection)
• raster_stats - Statistical analysis with histograms

• vector_info - Inspect metadata (CRS, geometry, attributes)
• vector_reproject ⚡ - CRS transformation (with reflection)
• vector_convert - Format migration (SHP ↔ GPKG ↔ GeoJSON)
• vector_clip - Spatial subsetting
• vector_buffer - Proximity analysis
• vector_simplify - Geometry simplification

• store_justification - Cache epistemic reasoning (used internally)
• Advisory prompts for CRS selection and resampling methods

⚡ = Reflection-enabled: These tools require methodological justification on first use, then cache for instant subsequent execution.

See TOOLS.md for complete documentation with examples and parameters.

# Run all tests
uv run pytest test/ -v

# With coverage
uv run pytest test/ --cov=src --cov-report=term-missing

Status: ✅ 72 passing tests including reflection system integration

Python-Native Stack (ADR-0017):

• Rasterio - Raster I/O and manipulation
• PyProj - CRS operations and transformations
• pyogrio - High-performance vector I/O (fiona fallback)
• Shapely - Geometry operations
• NumPy - Array operations and statistics
• Pydantic - Type-safe models with JSON schema

Key Design Decisions (26 ADRs guide development):

• ADR-0026: Reflection system and epistemic governance
• ADR-0017: Python-native over CLI shelling for performance
• ADR-0011: Explicit resampling required (prevents silent data corruption)
• ADR-0022: Workspace isolation for security

We welcome contributions! See CONTRIBUTING.md for:

• Development setup
• Code style guide (Ruff + mypy)
• Testing requirements (pytest + fixtures)
• ADR process

MIT License - see LICENSE for details.

• Built with FastMCP
• Powered by Rasterio and GDAL
• Inspired by the Model Context Protocol

Current Status: v1.1.1 - Phase 2 Complete ✅

• Reflection middleware operational
• Vector/raster tool parity achieved
• Cross-domain cache sharing validated (75% hit rates)

Next: Phase 3 - Workflow Intelligence (v2.0+)

• Formal workflow composition
• Multi-step orchestration
• Analysis pattern libraries

See Vision for the complete long-term roadmap.

Built with ❤️ for the geospatial AI community

Geospatial operations that think, not just execute.