A task-oriented LLM agent written in Crystal. Not a chatbot - runs as a process, completes the task, exits.
Design Philosophy
- Task-oriented - spawned with a goal, works until done
- Tool mastery - bash execution, file reading with structural awareness
- Local-first - runs Gemma 4 via Ollama, no cloud dependencies
- Tight loop - executes tools, feeds results back, iterates
Technical Stack
- Crystal - compiled performance, Ruby-like syntax
- Gemma 4 26B - local model with native tool calling
- Ollama - model serving and inference
- OpenTelemetry - observability built-in