Add MCP gateway: drop Lemonade into Claude Desktop, VS Code or Cursor in 30 seconds

[siavashhub]

What this does. Adds a single POST /mcp endpoint that speaks the Model Context Protocol (Streamable HTTP, spec 2025-06-18). Any MCP-aware host like GitHub Copilot (Agent mode), Claude Desktop, Cursor, Cline, Continue... can now call local Lemonade models as tools.

Why it matters. Agents running in a frontier model can now route the privacy-sensitive or high-volume parts of a task to local Lemonade models without leaving the conversation. Bulk classification, on-device transcription and image generation become free, private and offline, orchestrated by whatever host model the user prefers.

MVP scope. 4 tools:

• lemonade_list_models - discovery
• lemonade_chat - chat completion (forwards to any backend)
• lemonade_transcribe_audio - base64-in, transcript-out
• lemonade_generate_image - base64 PNG out

Embeddings, TTS and SSE streaming intentionally deferred (happy to add as follow-ups if merged.)

Try it (≈30 sec). Start lemond, then add to VS Code's

{
  "servers": {
    "lemonade": {
      "type": "http",
      "url": "http://localhost:13305/mcp"
    }
  }
}

Or use below within claude_desktop_config.json for Claude Desktop

"mcpServers": {
  "lemonade": {
    "command": "npx",
    "args": [
      "-y",
      "mcp-remote",
      "http://localhost:13305/mcp"
    ]
  }
}

Open Copilot Chat in Agent mode → the four lemonade_* tools appear in the picker.

GHCP Screenshot 1:

GHCP Screenshot 2:

Diff scope. Additive only, no breaking changes to existing REST endpoints, no new runtime deps. Roughly ~600 lines of C++ plus tests/docs.

Files added

• src/cpp/include/lemon/mcp_server.h - McpServer class declaration
• src/cpp/server/mcp_server.cpp - JSON-RPC 2.0 dispatcher + 4 tools
• test/server_mcp.py - integration tests (envelope behaviour, tools/list, live chat call)
• docs/api/mcp.md - endpoint reference with curl examples

Files modified

• src/cpp/server/server.cpp - instantiate McpServer next to OllamaApi; add /mcp to is_api_route so LEMONADE_API_KEY still gates it
• CMakeLists.txt - register mcp_server.cpp in SOURCES_CORE
• AGENTS.md- document/mcpas the third exception to the quad-prefix registration (alongside Anthropic/v1/messagesand Ollama/api/*`); MCP spec mandates a fixed single-URL shape
• docs/api/README.md - new row in the API table

Roadmap (separate follow-ups if there's appetite). Embeddings + reranking tools; TTS; SSE streaming for chat; an MCP client in Lemonade (would unlock Lemonade-calls-Lemonade and voice-driven agents over the Realtime API).

Tests. python test/server_mcp.py covers envelope behaviour (parse errors, method-not-found, batch, notifications) and a live chat round-trip. Local build clean on Windows.