Mantle

Mantle is a self-hosted AI brain that ingests files, mail and notes, builds a layered memory with a knowledge graph, and serves it to Claude, Cursor or any MCP client, with four access tiers and federation between brains.

Mantle is a self-hosted knowledge system built around a single idea: instead of switching to yet another assistant, a person or organisation keeps one brain on their own server and lets every AI client draw from it. Files, mail, voice notes and anything created in the companion app Jackdaw are ingested without tagging or schema work. Facts, entities, links and embeddings are extracted as content arrives, then organised into a layered memory model that runs from persona and journal through conversation digests and profile facts to a searchable index and an append-only store of originals. In practice, the brain is reached through several doorways. Claude, Codex, Cursor or any MCP client connects as a tool server, where a fast recall path lets an agent state its task in a line and receive the matching prompt or procedure the owner wrote, while grounded search returns facts, documents and tables with citations. Telegram, email, share links, team and client logins and an API cover the rest. A Toolsmith agent reads a service's API documentation and writes tested tools, with credentials kept in an encrypted vault, and heartbeats run scheduled routines. Ingestion covers email, Microsoft 365, Telegram, files, voice and subscribed calendars. Access control is a central design point. Four levels (admin, team, client and public) are enforced by the database, so answers match across the app and MCP. Content is treated as data rather than instructions to limit prompt injection. The whole stack runs on one Postgres instance covering vectors, graph, full-text search and queues, in a Docker-based deployment. Mantle suits developers, small teams, agencies with client portals and privacy-conscious organisations that want persistent context under their own control. Compared with hosted assistants or vector-database-based retrieval setups, it trades convenience for ownership and breadth. The stated ceiling is roughly 20,000 documents per brain, after which the recommended path is adding further brains that query each other through scoped federation grants.