OpenACM — Open Autonomous Agent

OpenACM is a self-hosted autonomous AI agent that runs on your PC. It controls your local environment, writes and executes code, navigates the web, and connects to any MCP server — all through a modern web dashboard.

No subscriptions. No cloud dependency. Your data stays local.

Created and maintained by Jeison Hernandez / JsonProductions.
If you use or build on OpenACM, a credit or a star goes a long way.

• Run commands & code — executes shell commands and stateful Python (Jupyter kernel)
• Browse the web — Playwright-powered browser automation: login, scrape, screenshot
• MCP Server support — connect to any Model Context Protocol server (unity-mcp, filesystem, custom tools, etc.)
• Multi-channel — chat via Web, Telegram, or Console, all sharing the same AI brain
• Skills system — define reusable Markdown-based skills the AI triggers automatically
• Sub-agents — spawn specialized agents that work in parallel on tasks
• RAG memory — ChromaDB long-term memory that persists across conversations
• Local intent router — hybrid local/cloud architecture that skips the LLM for simple commands (~5ms, no tokens spent)
• Loop trace debugger — inspect every iteration: context size, tool calls, LLM timing, truncations

• Python 3.12+
• Node.js 20+
• An API key from any supported LLM provider

git clone https://github.com/Json55Hdz/OpenACM.git
cd OpenACM
.\setup.bat       # first time: installs everything and launches OpenACM

Next time just run:

.\run.bat

git clone https://github.com/Json55Hdz/OpenACM.git
cd OpenACM
chmod +x setup.sh run.sh
./setup.sh        # first time: installs everything and launches OpenACM

Next time just run:

./run.sh

docker-compose up -d --build
docker logs openacm   # your dashboard token is printed here

Open http://localhost:47821, paste the token, done.

1. The console prints your Dashboard Token — copy it
2. Open http://localhost:47821
3. Paste the token to log in
4. Go to Configuration and add your LLM API key

Uses LiteLLM internally — any provider it supports works:

Connect to any MCP server from the MCP Servers dashboard page:

Once connected, the AI automatically sees and uses those tools.

OpenACM/
├── frontend/                   # React + Next.js dashboard
│   ├── app/                    # Page routes
│   ├── components/             # UI components
│   ├── hooks/                  # API and WebSocket hooks
│   └── stores/                 # Zustand state
├── src/openacm/
│   ├── core/
│   │   ├── brain.py            # Agentic loop + trace system
│   │   ├── llm_router.py       # LiteLLM interface + retries
│   │   ├── local_router.py     # Local intent classifier
│   │   ├── memory.py           # Conversation memory
│   │   └── rag.py              # Vector memory (ChromaDB)
│   ├── tools/
│   │   ├── mcp_client.py       # MCP server manager
│   │   ├── registry.py         # Tool registry
│   │   └── ...                 # Built-in tools
│   └── web/
│       └── server.py           # FastAPI server + WebSocket
├── skills/                     # Built-in skill definitions
├── config/                     # Local config (not committed)
├── setup.bat / setup.sh
└── run.bat / run.sh

OpenACM is fully self-hosted. The only outbound traffic is what you explicitly trigger:

• LLM API calls to the provider you configured
• Telegram/Discord messages if you connect those channels
• Browser requests when you ask it to visit a site

Everything else — conversations, API keys, files, memory — lives in data/ and config/ on your machine. Use Ollama for a fully offline setup.

Contributions are welcome. See CONTRIBUTING.md.

MIT — free to use, modify, and distribute.
Copyright (c) 2026 Jeison David Hernandez Pena (JsonProductions). All copies and derivatives must include the original copyright notice.