🔌 MCP (Model Context Protocol)
An open protocol that standardizes how AI applications (hosts/clients) connect to external tools, resources, and prompts through MCP servers. “USB-C for AI integrations.”
Concepts
- Architecture: host → client → server; JSON-RPC 2.0 messages; capability negotiation
- Transports: stdio (local) and Streamable HTTP (remote)
- Server primitives: tools (model-controlled actions), resources (app-controlled context), prompts (user-controlled templates)
- Client features: sampling, roots, elicitation
- Auth for remote servers: OAuth 2.1-based authorization
- Security: tool poisoning, prompt injection via tool output, confused deputy, least privilege, user consent for sensitive tools ⭐
- Ecosystem: official SDKs (TypeScript, Python, Java, Go, Kotlin, C#), MCP Inspector, registries
🧪 Labs (🟢 warm-up → 🟡 core → 🔴 hard → ⚫ boss)
- 🟢 Run an existing MCP server (filesystem/GitHub) with MCP Inspector
- 🟡 Orbit MCP server in Spring AI (tools + resources)
- 🔴 A read-only Go MCP server (official Go SDK) for run logs/metrics
- 🔴 Remote HTTP transport with OAuth + per-tool approval
- ⚫ An MCP client in orbit-tools: dynamic tool discovery from external servers with scopes
🧠 Cognitive tasks
- Threat model: tool poisoning, a confused deputy, injection via tool output
🛰️ Orbit integration
- Orbit is usable from Claude Desktop / Claude Code / IDEs as an MCP server
Go deeper
Resources
- modelcontextprotocol.io (spec, docs, quickstarts) ⭐ · SDK repos on GitHub · Spring AI MCP docs