ACR: local-first AI runtime with memory, skills, agents—no cloud needed

ACR (Adaptive Cognitive Runtime) is a Python CLI tool designed to run as a local-first AI orchestration layer on your own machine. It includes a real task engine, a hybrid-retrieval memory store, a skill registry with routing and validation, model and tool routing with safe-mode-aware permission checks, an agent planner and critic system, an operational dashboard with telemetry visualization, and an MCP server exposing ACR's memory, skill, web, and GitHub search tools to any MCP client—including direct integration with Claude Code and Codex CLI through project-scoped registration.

The runtime is entirely self-contained and relies on SQLite for persistence; it does not require a cloud account, API key, or telemetry to operate. Users can install via `pip install acr-runtime` (or `uv tool install acr-runtime`) and then run `acr db upgrade` to initialize the local database schema. The tool supports pointing to local Ollama models for real inference or using a zero-config mock provider indefinitely. An operational dashboard is available at http://127.0.0.1:8765 when run with `acr dashboard serve`.

For development use, ACR can be installed from a source checkout using `uv sync`, which installs dependencies and the local package into a virtual environment. The project includes development workflows for testing (`uv run pytest`), linting (`uv run ruff check .`), formatting (`uv run ruff format .`), and type checking (`uv run pyright`). The full implementation state is documented in docs/ARCHITECTURE.md, which serves as the authoritative source for what is actually built and why.

The project is MIT licensed and includes security reporting procedures in SECURITY.md. Bug reports, questions, and feature requests are handled via GitHub Issues. Claude Code and Codex CLI projects can register ACR's MCP server without manual setup; Claude Code prompts for approval on first use, while Codex CLI requires the project to be explicitly marked trusted via `codex trust` or configuration before loading project-scoped tools, ensuring that cloning the repo cannot silently launch any functionality without user consent.

ACR addresses a core concern in the AI development community: the desire to run AI workloads without outsourcing data or inference to cloud providers. By bundling persistent memory, task orchestration, skill routing, and model management all on SQLite—a zero-dependency, file-based database—the project eliminates infrastructure friction for local-first workflows. The choice to support both local Ollama models and a mock provider reflects a design philosophy that accommodates users at different stages: those experimenting without hardware can prototype with the mock provider, and those with local inference capacity can point ACR at their own models.

The integration into Claude Code and Codex CLI via the Model Context Protocol (MCP) is significant because it treats ACR not as a standalone tool but as a primitive that other applications can compose. By registering MCP servers at the project level, the runtime avoids silent privilege escalation; both Claude Code and Codex CLI require explicit user consent (approval prompts or explicit trust marking) before loading any registered tools. This design acknowledges the trust boundary between developer intent and ambient automation.

The project's architecture—documented in ARCHITECTURE.md as the single source of truth for implementation state—signals a commitment to clarity about what is actually built versus aspirational. By tying all specification and status to living documentation rather than release notes, the maintainers reduce the risk of claims drifting from reality, a common failure mode in open-source announcements.