CantusCompose LLM agents like polyphony

A teaching-oriented, Colab-first framework for building agent harnesses from two protocols — Skill and Memory.

Two protocols, one loop

Skills are callable behaviours; Memory holds state. An Agent runs the read–act–observe loop over them, recording every step in an EventStream.

Local or cloud models

Run Gemma on Colab, MLX on Apple Silicon, or any of eight providers — OpenAI, Anthropic, Google, Groq, NVIDIA, Ollama, MLX, and a local OpenAI-compatible server.

Serve and observe

Expose a registry over HTTP with `cantus serve`, connect LINE, Telegram, Discord, or Google Chat, and watch sessions live in the `cantus tui` dashboard.

Built for the classroom

Designed so students write a skill, swap a provider, and inspect what the agent did — without leaving a notebook.