Local-first AI runtime

ACR

Adaptive Cognitive Runtime

Persistent memory, skills, and task orchestration that run entirely on your own machine. No cloud account, API key, or telemetry required to try it.

$pip install acr-runtime

$acr db upgrade && acr doctor

$acr run "say hello"

GitHub PyPI Docs

Manifest

Task engine: Create, execute, and track work end to end, with full telemetry on every run.
Memory: Hybrid keyword and metadata retrieval over SQLite FTS5, with confidence-scored write control before anything becomes a trusted fact.
Skills: A registry with routing, validation, and versioned evolution, promoted only on evidence.
Model routing: An escalation ladder from free and local to paid and cloud. Cheapest qualifying model first, always.
Agents: A planner and critic loop, spawned only when the cost and risk estimate says it is worth it.
Dashboard: A local, read-only view into everything ACR has done, with a real-telemetry visualization.
MCP server: Exposes ACR's own memory and skill tools to Claude Code, Codex CLI, or any MCP client.

Why local-first

Nothing leaves your machine unless you ask it to. ACR defaults to a zero-config mock provider, so there is nothing to configure just to try it. Point it at your own local Ollama when you want real inference, or a cloud key you set yourself. It never assumes you are online, and it never assumes you are paying for anything.

MIT licensed. 95% test coverage. Built to run on hardware you already own.

ACR

Get started