The reason to run this is not that it connects to Slack and Figma — it is that connecting to both does not cost you the context window. Tools are embedded at startup, each request expands into a few action phrases, and cosine similarity returns the handful worth showing the model. The security posture is worth reading too: tokens are AES-256 encrypted at rest and never exposed to the model, every tool call is written to an audit log, and memory refuses to auto-save anything that looks like a password or key.
A self-hosted AI operations platform that presents itself to MCP clients as a handful of meta-tools. Instead of exposing every integration's schema on every call, a local sentence-transformers model ranks tools by intent and returns only the few that match. Around that sit persistent vector memory, an automation scheduler, multi-agent flows and a ReAct execution loop.
- Discovery and execution through four hub tools: list the connected integrations, find tools by semantic intent, browse one provider's tools, and call a tool by exact name — `hub.integrations.list`, `hub.tools.search`, `hub.tools.list`, `hub.tools.call`
- Integrations with their own tool sets: Teamwork over OAuth for projects, tasks, dependencies and time tracking; Slack for channels, DMs and canvases; Miro for boards and items; Figma for CSS extraction and layout trees; Google Calendar; Binance for market data and spot trading; Telegram over MTProto and a Bot API path
- Persistent memory with hybrid retrieval — FTS5 keywords plus vector similarity — typed items with automatic expiry, secret detection that blocks saving credentials, and a context pack call that returns only what is relevant to the current task — `memory.search`, `memory.upsert`, `memory.summarize_context`
- A scheduler with interval, cron and event triggers, whose payload can be a notification, a specific tool call, or a full agent run with tool access
- Multi-agent flows: graphs of agents with their own prompts, tool policies and mounted skills, delegating to each other
- Web search built in — search, page fetch, and a combined call that opens the top results — `web_search.search`, `web_search.open`, `web_search.full_search`
Docker for the recommended path, or Python with `uvicorn` locally. Five required settings before first run: `ADMIN_EMAIL`, `ADMIN_PASSWORD`, `JWT_SECRET`, `TOKENS_ENCRYPTION_KEY` and `BASE_URL`. Clients connect to the /mcp endpoint with a Personal Access Token as a bearer header; ChatGPT can discover the endpoints over OAuth instead. Each integration needs its own OAuth client id and secret, set in the environment or through the admin UI. Storage is SQLite with automatic migrations. The published production image is CPU-only by design.
One command plus a key — docker-compose up -d, then supply credentials