Google Maps for AI agents
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
**Stop rebuilding place search, routing, and location workflows for every AI app.**
@cablate/mcp-google-map turns Google Maps Platform into 18 read-only agent tools and three focused Skills. An agent can find real places, verify routes, compare options, build practical itineraries, or audit local search visibility—through MCP or a standalone CLI.
You choose the integration model: Codex or Claude Code Plugin when you want Skills without MCP setup, stdio for desktop MCP clients, or Streamable HTTP for shared and remote deployments.
English · 繁體中文
Why use it?
Giving an agent a raw Maps API is only the beginning. Useful answers often require several dependent calls: resolve a place, preserve its identity, search around it, check opening details, calculate travel time, and explain what was actually verified. This project packages that work into one consistent interface.
- **Move from lookup to outcome.** Atomic tools handle geocoding, place details, directions, weather, air quality, and maps. Composite tools explore areas, compare candidates, optimize stops, and measure local rankings.
- **Use the same capabilities with or without MCP.** The standalone exec CLI works with Agent Skills and automation; the MCP server exposes the same 18 tools over stdio or HTTP.
- **Give agents workflow guidance, not just function names.** Three Skills cover general map research, evidence-backed travel planning, and local SEO. Codex loads their full instructions only when a request matches.
- **Keep deployment and credentials under your control.** Run locally or self-host. API keys remain in your environment or request headers, with per-session isolation for HTTP deployments.
- **Start narrow and grow later.** Register only the tools you need with GOOGLE_MAPS_ENABLED_TOOLS, or use the full catalog.
What can an agent do?
| Outcome | How the project helps |
|---|---|
| Find and evaluate real places | Natural-language and nearby search, place details, ratings, hours, reviews, and distance comparison |
| Build a trip that works geographically | Candidate discovery, along-route stops, travel-time checks, multi-stop optimization, weather, and static maps |
| Research a neighborhood | Multi-category exploration plus targeted distance, elevation, timezone, weather, and air-quality checks |
| Plan field work or deliveries | Route matrices and optimized stop ordering for up to 25 stops |
| Audit local search visibility | Geographic grid ranking, competitor discovery, ARP, ATRP, and SoLV metrics |
| Enrich location data | Single or batch geocoding, reverse geocoding, and structured JSON output |
These are data and planning tools, not guarantees of safety, accessibility, opening status, or ranking outcomes. Applications displaying Places reviews, photos, or AI summaries must follow the content attribution and storage guidance.
Choose your integration
| Use | Best for | What runs |
|---|---|---|
| **Codex Plugin** | Asking Codex map, travel, or local SEO questions without MCP configuration | A matching Skill loads on demand and calls the standalone CLI |
| **Claude Code Plugin** | Installing the same three Skills from a Claude marketplace | Skills are namespaced under mcp-google-map and call the standalone CLI |
| **Standalone CLI** | Scripts, automation, and other Skill-compatible agents | One stateless command returns JSON |
| **MCP stdio** | Claude Desktop, Cursor, VS Code, and other local MCP clients | The client starts a local MCP process |
| **Streamable HTTP** | Multi-session, containerized, LAN, or remote access | A self-hosted server exposes /mcp |