Reflection MCP Server

Part of the LXD MCP Suite — a cohesive set of MCP servers for learning experience design (coaching, Kanban, stories, and optional LLM adapters).

What it is

Lightweight reflection MCP server (stdio) that detects available providers and stores short local memories.

Why it helps

Optional tailoring/validation for other servers; stays small and safe. Works fully offline with local memory only.

Lightweight reflection and differential diagnosis MCP server.

• Detects provider from environment/.env (OpenAI, Anthropic, Gemini, Ollama) and uses a lightweight local model if no network provider is available.

• Stores short, bounded memories per key in .local_context/reflections/<key>.jsonl.

• Exposes MCP tools over stdio:

  • reflection_handshake(user_key, name)

  • reflect(key, input)

  • ask(key, question)

  • note(key, note)

  • memories(key, limit?)

  • summarize(key)

Quickstart

Register with an MCP client (example)

• Claude Desktop (config snippet):

Environment variables

• OpenAI: OPENAI_API_KEY, OPENAI_BASE_URL (optional), OPENAI_MODEL (default: gpt-4o-mini)

• Anthropic: ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL (optional), ANTHROPIC_MODEL (default: claude-3-haiku-20240307)

• Gemini: GOOGLE_API_KEY, GEMINI_BASE_URL (optional), GEMINI_MODEL (default: gemini-1.5-flash)

• Ollama: OLLAMA_BASE_URL or OLLAMA_HOST, OLLAMA_MODEL (default: llama3.1:8b-instruct)

If no provider key is found or requests fail, the server falls back to a local lightweight reflector.

File layout

• reflection_mcp/mcp_server.py: MCP stdio server

• reflection_mcp/provider.py: provider detection + HTTP client

• utils/reflection_memory.py: shared local memory store (JSONL)

Install (local PATH)

Run at Login

macOS (launchd)

Linux (systemd user)

License

Proprietary/internal by default. Add a license if open-sourcing.

Internal Use Only — not licensed for external distribution or hosting.