cesium mcp

[!IMPORTANT]

📣 作者正在求职 · 北京

AI 应用开发 · Agent 应用开发 · AI 产品全栈

我是高鹏彬，有约 6 年软件开发经验。如果你们团队正在招人，欢迎联系，也感谢帮忙内推或转发！

📄 查看简历 PDF　·　✉️ 联系我

查看我的项目与个人介绍 → · 邮箱：1804287165@qq.com

[!TIP] 📣 Built with Cesium MCP? / 谁在使用 Cesium MCP？

Share your project, screenshots, or feedback—work in progress is welcome! / 欢迎分享项目、截图和使用反馈，开发中的作品也欢迎！

Share / 分享 → Issue #44 · Discussion / 讨论区

---

Demo

https://github.com/user-attachments/assets/8a40565a-fcdd-47bf-ae67-bc870611c908

Packages & Entry Points

Which one? Personal project or quick try → browser-agent. Let a compatible browser agent discover page-local Cesium tools → WebMCP. Existing web app embedding an AI assistant → bridge + your own function calling. Calling from Claude Desktop / Cursor / Dify → MCP runtime.

Architecture

flowchart LR
  subgraph clients ["AI Drivers (pick one)"]
    BA["Browser Agent\n(in the same page)"]
    WM["WebMCP Agent\n(browser-provided)"]
    FC["Your web app\nfunction calling"]
    MCP["Claude / Cursor / Dify\nvia MCP runtime"]
  end

  CONTRACTS["cesium-mcp-contracts\ntool definitions"]
  WEBMCP["cesium-mcp-webmcp\nnative adapter"]

  subgraph core ["cesium-mcp-bridge (browser)"]
    B["60+ tools\nprotocol-agnostic dispatcher"]
    C["CesiumJS Viewer"]
  end

  CONTRACTS -.-> BA
  CONTRACTS -.-> WEBMCP
  BA -- "in-page call" --> B
  WM -- "document.modelContext" --> WEBMCP
  WEBMCP --> B
  FC -- "in-page call" --> B
  MCP -- "WebSocket / JSON-RPC" --> B
  B --> C

  style clients fill:#1e293b,stroke:#528bff,color:#e2e8f0
  style core fill:#1e293b,stroke:#12B76A,color:#e2e8f0

The bridge remains the execution core, while contracts and protocol adapters stay separate. Pick whichever driver matches your scenario — they all reach the same Cesium command layer. On WebMCP-capable browsers, cesium-mcp-webmcp can expose 61 browser-safe commands in 12 selectable toolsets through document.modelContext without adding an MCP transport or backend server.

Relationship to the CesiumGS AI ecosystem

CesiumGS's newer AI work is split between cesiumjs-ai-starter-app, a deployable application template, and cesiumjs-skills, development-time guidance for coding agents. The earlier cesium-ai-integrations repository contains the first-generation experiments and community contributions that helped explore this space.

cesium-mcp is an independent runtime and integration toolkit, not a continuation of the earlier WebSocket-only reference architecture. Its reusable Bridge and shared contracts work unchanged across browser-only function calling, native WebMCP, standard MCP over stdio/HTTP, and embedded desktop shells. A local WebSocket bridge is used only when an external MCP host needs to reach a live browser Viewer; it is not required for the hosted demo or page-local integrations.

The project author was an early contributor to CesiumGS/cesium-ai-integrations, contributing the Imagery server, Terrain server, and unified MCP Gateway. Those experiments informed this project's multi-protocol architecture, while the implementation, release lifecycle, and roadmap remain independent.

Quick Start

Path 0 — Try in 30 seconds (browser agent, recommended)

Open the live demo and ask—the hosted model is ready without a browser API key:

"Fly to the Eiffel Tower and drop a red marker"

Fork the examples/browser-agent folder to deploy your own.

Path 1 — Expose Cesium tools through WebMCP (Chrome 149+ experimental)

The browser-agent example automatically registers all 61 browser-safe page tools when document.modelContext is available. Its built-in chat uses automatic toolset routing to keep each normal request at 20 tools or fewer, while still offering explicit core, single-toolset, and all-61 modes:

npm run build -w packages/cesium-mcp-bridge
npm run build -w packages/cesium-mcp-webmcp
npx serve . -l 4173

Open http://localhost:4173/examples/browser-agent/, click Start, then inspect or execute the tools in DevTools → Application → WebMCP. Enable #enable-webmcp-testing and #devtools-webmcp-support in chrome://flags for local testing.

Application developers install the adapter separately. End users only open the integrated website; they do not install npm packages or run an MCP server.

npm install cesium-mcp-webmcp

import { registerCesiumViewerWebMcp } from 'cesium-mcp-webmcp/viewer'

const registration = await registerCesiumViewerWebMcp(viewer, {
  toolsets: 'all',
  excludeTools: ['geocode'], // add your own browser geocoder to expose this tool
})

// Later, if the page is unmounted:
registration.unregister()

See the WebMCP adapter API for custom integrations. For a complete npm + Vite application, start from the WebMCP integration example.

Path 2 — Embed in your own web app (function calling)

npm install cesium-mcp-bridge

import { CesiumBridge } from 'cesium-mcp-bridge';

const bridge = new CesiumBridge(viewer);
// Then: send the bridge's tool schema to any LLM that supports function/tool calling,
// route the model's tool calls to bridge.execute(name, params).

See examples/browser-agent/index.html for a complete loop with OpenAI-compatible APIs.

Path 3 — Use from Claude Desktop / Cursor / Dify (MCP)

Ordinary MCP users need only the Runtime package. It includes the browser Bridge bundle and a built-in Viewer at http://localhost:9100/; install cesium-mcp-bridge separately only when integrating a custom page.

# Stable channel — npm latest, MCP SDK v2
npx -y cesium-mcp-runtime

# HTTP mode
npx -y cesium-mcp-runtime --transport http --port 3000

The stable release serves existing MCP 2025-11-25 clients and the new 2026-07-28 protocol from the same stdio/HTTP entry. It uses the stable TypeScript SDK v2 and passes the official server-stateless conformance scenario (28/28).

MCP client config:

{
  "mcpServers": {
    "cesium": {
      "command": "npx",
      "args": ["-y", "cesium-mcp-runtime"]
    }
  }
}

62 Available Command Tools

Tools are organized into 12 toolsets. Default mode enables 4 core toolsets (30 tools). Set CESIUM_TOOLSETS=all for everything, or let the AI discover and activate toolsets dynamically at runtime.

Canonical contracts: Tool descriptions default to English; set CESIUM_LOCALE=zh-CN for Chinese. Titles, behavior annotations, localized descriptions, defaults, input validation, MCP output schemas, and structured results all come from the shared JSON Schemas in cesium-mcp-contracts. Text content remains available for older clients.

Examples

See examples/minimal/ for a complete working demo.

Development

git clone https://github.com/gaopengbin/cesium-mcp.git
cd cesium-mcp
npm install
npm run build
npm test
npm run test:contracts
npm run test:schema-compat
npm run test:routing
npm run test:model-tools
npm run eval:model-tools
npm run test:e2e:packed

test:contracts is the focused parity gate for MCP Runtime metadata, WebMCP registration, Function Calling definitions, provider Schema portability, and the 60-tool Bridge Executor Registry. Run test:schema-compat directly for actionable OpenAI, Azure, VS Code MCP, and WebMCP Schema diagnostics. test:routing evaluates bilingual and multi-intent Browser Agent requests across all 12 toolsets, checking required-tool recall and the 20-tool automatic-routing budget. test:model-tools verifies the provider-neutral multi-turn scoring harness. eval:model-tools performs a no-network routing preflight by default; add an explicit provider and --live to measure real tool choice, argument validity, and required-tool completion. See Model Tool Evaluation. test:e2e:packed builds npm tarballs, installs them in a clean temporary project, opens the real Cesium Viewer, and verifies a Runtime-WebSocket-Bridge command round trip.

Version Policy

Version format: {CesiumMajor}.{CesiumMinor}.{MCPPatch}

Official CesiumJS releases are reviewed before the compatibility baseline is bumped; the project does not automatically claim support for a newer release without Bridge verification.

Related Projects

• mapbox-mcp — AI control for Mapbox GL JS
• openlayers-mcp — AI control for OpenLayers

Community

This project recognizes LINUX DO as a community for open-source exchange, technical discussion, and developer feedback.

Star History

License

MIT