ACR: Adaptive Cognitive Runtime

acr-runtime.netlify.app

A local-first AI orchestration runtime that runs entirely on your own machine. Persistent memory, a skill system, task/agent orchestration, and multi-provider model routing, all on SQLite, with no cloud account, API key, or telemetry required to try it. Point it at your own local Ollama models when you want real inference, or keep it on the zero-config mock provider indefinitely; ACR never assumes you're online or paying for anything.

pip install acr-runtime
acr db upgrade && acr doctor
acr run "say hello"

Full specification: ACR_MASTER_SYSTEM_PROMPT.md. Current implementation state: docs/ARCHITECTURE.md, the single source of truth for what's actually built, so nothing here duplicates that status and drifts stale.

What's here

A Python CLI (acr) built around a real task engine, a hybrid-retrieval memory store, a skill registry with routing/validation/evolution, model and tool routing with safe-mode-aware permission checks, an agent planner/critic/topology system, an operational dashboard with a real-telemetry visualization, and an MCP server exposing ACR's memory, skill, web, and GitHub search tools to any MCP client, including project-scoped registration for both Claude Code and Codex CLI out of the box. See docs/ARCHITECTURE.md for the full breakdown of what's implemented, how, and why.

Quick start

Install from PyPI:

pip install acr-runtime   # or: uv tool install acr-runtime
acr db upgrade            # create the local SQLite schema, no repo checkout needed
acr doctor

Or from a source checkout (for development; see CONTRIBUTING.md):

uv sync                 # install deps + local package into .venv
cp .env.example .env    # local dev data dir (repo-local ./data, gitignored)
uv run alembic upgrade head
uv run acr doctor

acr run "say hello"          # create + execute a task end to end
acr dashboard serve          # operational dashboard: http://127.0.0.1:8765
acr mcp serve                # MCP server (stdio by default)

(prefix with uv run if working from a source checkout instead of a pip/uv tool install)

Using ACR from Claude Code or Codex CLI

This repo ships project-scoped MCP server registration for both Claude Code and Codex CLI, so opening it in either offers ACR's memory/skill/web/GitHub search tools and task execution as MCP tools directly. Claude Code prompts for approval the first time it opens the project; Codex CLI only loads project-scoped config for a project you've marked trusted (codex trust prompt, or trust_level = "trusted" under [projects."<path>"] in your own ~/.codex/config.toml); either way, cloning the repo can't silently launch anything without you consenting.

Development

See CONTRIBUTING.md for the full workflow. Short version:

uv run pytest            # tests
uv run ruff check .      # lint
uv run ruff format .     # format
uv run pyright           # type check

License, security, support

MIT licensed. See SECURITY.md to report a vulnerability privately. For bugs, questions, or feature requests, use GitHub Issues.