An open protocol (originally from Anthropic) that standardizes how AI agents connect to external tools and data sources. MCP defines a client-server architecture where agents (clients) invoke capabilities exposed by tool servers through a standard interface, similar to how USB-C standardizes device connectivity. It is closely related to Tool calling.
Adoption in Geospatial
As of early 2026, MCP is rapidly becoming the standard connector between AI agents and GIS tools. Multiple MCP servers exist for PostGIS, QGIS, GeoServer, OpenStreetMap, and geocoding services. The GIS MCP server (gis-mcp) reaches ~4,300 weekly downloads on PyPI (early 2026). (PulseMCP)
- CARTO MCP Server: exposes Geospatial AI workflows as tools callable by any MCP-compliant agent (Gemini, ChatGPT, Claude).
- Google Maps Platform MCP: geocoding, routing, and place search APIs accessible to LLMs.
Broader Ecosystem
MCP enables the agentic orchestration pattern: LLMs reason over problems while specialized tools handle domain execution rather than trying to embed all domain knowledge in model weights.
This decoupling is visible in CARTO’s architecture, Google’s Geospatial Reasoning framework, and Esri’s modular AI skills system.
Status as of early 2026
MCP flaws are becoming more and more problematic for complex agent orchestration and sophisticated agent harnesses (eating up the context window of LLMs, creating potential confusion, etc.).
More “naive” tool discovery like calling simple hierarchical tool lists via bash seems to be ass effective in some environments, and can easily be adapted to many different MCP use cases.
The real underlying problem might have also been the hype around the protocol that led to a rapid inflation of available MCPs and hence their misuse (MCP is Dead; Long Live MCP!)
Loading comments...