Ten packages. One capable agent.
Each piece is small, typed, and tested — and composes into the agentic-OS kernel. Nothing is a black box.
Any model
Anthropic, OpenAI, OpenRouter, Ollama, or any OpenAI-compatible endpoint.
Signed thinking-impressions
Every thinking window mints a runtime-signed proof — the trust anchor for ads.
Model Context Protocol
A first-class MCP client; bring any server's tools into the agent.
Claude + Codex skills
Loads Claude SKILL.md and Codex AGENTS.md into one skill system.
Lifecycle hooks
Gate, mutate, or extend every step: SessionStart, Pre/PostToolUse, Stop, Thinking.
Teams of agents
Durable task store, atomic checkout, budgets — run for hours or days.
Cross-session memory
A file-backed memory index plus recall, persistent across sessions.
Guardrails
Secret redaction, permission policy, path sandbox, audit trail.
A kernel for autonomous work.
The kernel wires providers, tools, skills, hooks, memory, and orchestration into one runtime you can run for minutes or for days — locally, on your keys.