Skip to main content
Glama
Sign In
ACR MCP Server

ACR MCP Server

by hbkdad
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

ACR: Adaptive Cognitive Runtime

CI PyPI License: MIT

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.

Related MCP server: WEATHGARDS

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.

If ACR is useful to you, consider supporting its development: ko-fi.com/hbkcustomsinc.

Available Tools

6 tools
browser_fetchB

Render a URL in a real headless browser and return its visible text, including JS-rendered content web_fetch can't see.

Inputs: url (required), max_chars. Returns: object

github_searchC

Search GitHub issues and pull requests (read-only).

Inputs: query (required), limit. Returns: result

memory_searchC

Keyword search over ACR's local memory store.

Inputs: query (required), limit. Returns: result

run_taskB

Run an objective through ACR's task engine. Uses the zero-config mock provider unless min_quality_tier is raised (or Settings.default_min_quality_tier is configured) to prefer a configured Ollama/cloud provider instead.

Inputs: objective (required), min_quality_tier. Returns: object

skill_searchB

Keyword search over ACR's local skill registry.

Inputs: query (required), limit. Returns: result

web_fetchC

Fetch a URL over HTTP(S) and return its extracted text.

Inputs: url (required), max_chars. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    B3.2/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool targets a distinct domain: memory search, skill search, web fetching (two variants), GitHub search, and task execution. No overlaps exist.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern in snake_case (e.g., memory_search, web_fetch, run_task), with clear prefixes for related operations.

    Tool Count5/5

    Six tools provide a focused set covering search, fetch, and task execution without being excessive or minimal.

    Completeness2/5

    Significant gaps exist: memory and skill stores only support search (no create/update/delete), and run_task lacks status or result retrieval, limiting the workflow.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Resources

    Related MCP Connectors

    Related MCP Servers