mudler/LocalAGI

LocalAGI – Self‑hosted AI agents with a web UI

What it is

• A Go‑based platform that lets you run AI agents entirely on your own machine (CPU or GPU).
• Agents are configured through a browser‑based dashboard – no Python code or cloud keys are required.
• Each agent exposes an OpenAI‑compatible REST endpoint, so existing tools that talk to the OpenAI API can be pointed at LocalAGI.

Key capabilities (as described in the README)

Feature	What you get
No‑code agent creation	Build and edit agents via the Web UI; enable/disable reusable Skills (tool‑like extensions) without writing code.
Team‑work	Define multiple agents that can cooperate on a single prompt, forming “agent teams”.
Connectors	Ready‑made integrations for Discord, Slack, Telegram, GitHub Issues, IRC (and a generic MCP connector).
REST API	Full OpenAI‑style /v1/chat/completions endpoint plus extra routes for managing agents, skills, and knowledge bases.
RAG / Knowledge base	Built‑in short‑ and long‑term memory using the LocalRecall library; you can upload files, create collections, and run semantic search.
Planning & reasoning	Agents can plan steps and adapt their behaviour based on the context they receive.
Periodic tasks	Cron‑like scheduling for recurring actions.
Multimodal	Supports vision, text and image generation models when run on a GPU.
Custom actions (Go)	Drop‑in Go files (or inline snippets) that define new actions; the system interprets them at runtime, no compilation needed.
Observability	Real‑time status panels in the UI and detailed logs via the API.
Model flexibility	Works with any GGUF/GGML model served by the companion LocalAI server; you can also point it at any OpenAI‑compatible endpoint via LOCALAGI_LLM_API_URL.

How you run it

• The project ships Docker‑Compose files for several hardware setups (CPU‑only, NVIDIA, Intel Arc, AMD).
• Example command for a CPU‑only launch:

  git clone https://github.com/mudler/LocalAGI && cd LocalAGI
  docker compose up
  

• For GPUs you swap in the appropriate compose file (docker-compose.nvidia.yaml, etc.) and optionally set MODEL_NAME, MULTIMODAL_MODEL, IMAGE_MODEL environment variables to pick the desired models (e.g., gemma-3-12b-it).
• After the containers start, open http://localhost:8080 to access the dashboard.

Extending the platform

• Skills – reusable tool definitions stored under a skills/ folder; they can be edited in the UI or version‑controlled via Git.
• Custom actions – place Go source files in a directory pointed to by LOCALAGI_CUSTOM_ACTIONS_DIR. Each file must expose three functions (Run, Definition, RequiredFields). The system loads them automatically and makes them selectable in the UI.
• Library mode – The core can be imported as a Go library, allowing developers to programmatically create agents, pools of agents, and custom connectors.

Why you might care

• Privacy – All data stays on your hardware; there are no external API keys unless you deliberately point the system at a cloud endpoint.
• Local‑first AI – Useful for developers, hobbyists, or organisations that want to experiment with autonomous agents without paying for cloud inference.
• Plug‑and‑play – Because it mimics the OpenAI API, existing bots, scripts, or front‑ends can be switched to LocalAGI with minimal changes.

---

Bottom line: LocalAGI is a full‑stack, self‑hosted AI‑agent framework that combines a web UI, OpenAI‑compatible APIs, RAG, multimodal support, and extensible Go‑based actions, all runnable on consumer‑grade hardware via Docker.