Labsco
MCP SERVER

MyAITwin MCP

by LutoLabs ↗

19 tools for a personal RAG database you build from inside chat — knowledge and voice stored separately, every result cited.

Vector Stores & RAG Retrieval
Summary
Knowledge and voice are stored as different things.

That separation is the design decision worth understanding: a meeting transcript is what you know, your email style is how you say it, and a good follow-up email needs both. `search_for_creation` returns them as separate result sets for exactly that reason, which a single similarity search cannot do.

What it is

A hosted MCP server holding a personal RAG database with semantic search that you shape from conversation: you define the types and tags, store knowledge and voice separately, and retrieve either with the source cited. Two layers sit behind it — Postgres via Supabase for structured records with provenance, Pinecone for the vector embeddings.

What you get
  • Storing (5) — `add_knowledge`, `add_voice_note`, `add_document` with automatic chunking, `add_from_url`, `add_reference_record`
  • Retrieving (5) — `search_twin` across all knowledge, `search_for_creation` which returns skills and knowledge as separate result sets, `get_by_type`, `get_by_tag`, `list_recent`
  • Understanding (4) — `get_schema` for your types and item counts, `get_sources`, `find_patterns` for recurring themes, `synthesise` across items on a topic
  • Managing (5) — `get_welcome`, `update_knowledge`, `add_schema_type`, `update_schema_type`, `delete_knowledge`
  • Of the 19: 10 read-only, 8 write, and 1 destructive — `delete_knowledge`. Every tool carries `title`, `readOnlyHint` and `destructiveHint` annotations
  • Every result is cited with source and date and tagged personal, organisational or external
Requirements

An account at https://myaitwin.lutolearn.com, then either the personal MCP URL from `/create` pasted into your client as a custom connector, or the canonical endpoint https://myaitwin.lutolearn.com/mcp. Transport is Streamable HTTP; authentication is OAuth 2.1 with PKCE (S256) and dynamic client registration, with bearer tokens hashed at rest. A client with MCP capability is needed — the README names Claude Pro, Claude Team and ChatGPT Pro. Each account lives in its own namespace with rate limiting per tenant and an append-only audit log on every tool call. Free during early access.