How mrgoonie/human-mcp works
Project type and purpose
• This is a standalone backend server (Human MCP Server v2.10.0) implemented in TypeScript/Node.js (supports Bun), exposing both stdio and HTTP/SSE transports.
• It implements the “Model Context Protocol” (MCP), hosting 27 multimodal “tools” (Eyes, Hands, Mouth, Brain) for AI agents (Claude Desktop, OpenCode, etc.) to call for vision, document processing, image/video generation & editing, speech generation, and advanced reasoning.
File structure
• Top-level “src” folder:
• src/server.ts – entry point that boots transports and the MCP core
• src/transports – HTTP server (file‐upload, /mcp API, SSE routes) and stdio transport
• src/tools – tool implementations grouped into “brain”, “eyes”, “hands”, “mouth” subfolders
• src/utils, src/types, src/prompts – shared utilities, type definitions, built-in prompt templates
• Config & scripts: Dockerfile, docker-compose.yaml, .env.example, Cloudflare R2 integration, GitHub Actions, Claude & OpenCode config scaffolding.
• Tests: unit, integration, e2e suites against HTTP and stdio transports.
README for additional insights
• Architecture section already outlines a 4×layer view:
AI Agent ←→ Human MCP Server ←→ Google Gemini & processing libs (Playwright, Jimp, rmbg, ffmpeg)
• Transport types and Cloudflare R2 for file uploads.
• Detailed MCP Tool breakdown reinforces the grouping of capabilities.
Detailed coverage
• Environment/config layer separately (port config, API key injection)
• Test harness interactions (tests/integration hitting HTTP API) if desired for clarity
• Don’t crowd—group tools by capability and annotate counts (4 Eyes, 16 Hands, 4 Mouth, 3 Brain)