Not a toy. Not a wrapper. A runtime built for real workloads.
🧠
3-layer persistent memory
Master context, distilled long-term facts, and short-term episodic memory — all in SQLite. Agents start sessions already knowing what matters.
🔀
Multi-model routing
Claude, DeepSeek, GPT-4o, Gemini, Llama — swap models per turn, per session, or based on confidence score. No vendor lock-in.
🛠️
Rich tool execution
Bash, file I/O, web search, MCP servers, custom tools — up to 10 concurrent tool calls per turn with full approval controls.
📚
SkillBank learning
Successful task patterns are automatically captured and injected into future runs. Agents get better at your specific workflows over time.
⚡
Auto-summarization
Sessions auto-compress at 70% token pressure, preserving recent context. Runs never die from context overflow.
🔬
OMLS training pipeline
Opportunistic RL training from session trajectories. Fine-tune a LoRA adapter during idle hours, host it on Together AI, and route to it automatically.