RIGEL
An open-source engine for building AI assistants and agents: several language-model backends, tools, memory, voice, and a way for the rest of the desktop to talk to it.
- When
- 2021 to now
- Built with
- Python, Ollama, Groq, llama.cpp, D-Bus, MCP, ChromaDB
RIGEL is the project I've carried longest. It is an engine, not a chat app: the part that sits between language models and everything that wants to use one.
What the engine does
- Runs on more than one model. Local models through Ollama or llama.cpp, or hosted ones through Groq, behind one interface.
- Acts, not only answers. It can decide to use tools, through the Model Context Protocol, and run tasks made of several steps, including driving a browser.
- Remembers and looks things up. Conversation memory, and retrieval over your own documents.
- Speaks and listens. Speech to text, text to speech, and voice cloning.
- Is part of the desktop. It runs as a D-Bus service on Linux, so any program or shell script can ask it something. There is an HTTP API as well.
It is released under the AGPL, and it is meant to be taken apart: build your own assistant on it, or use one piece.
Where it came from
The first RIGEL was the first "AI" program I ever wrote. I was 16. The interface is pretty bad, and the description I wrote of how it works is completely wrong. But it still works.
It came back as the assistant for ztOS 2.0 on the Stargate, and then as this engine. It now also runs inside my shell and on my phone's home screen.