Self-hosted multi-agent AI platform with voice, visual workflows, and tool integration.
Website · Docs · Quick Start
Define multiple AI agents, each with its own LLM, memory, and tools. Chain them into multi-step workflows. Access via voice, Telegram, webhooks, or cron. Manage it all from a visual admin panel.
Your server, your data, your rules.
curl -fsSL https://raw.githubusercontent.com/achetronic/magec/master/scripts/install.sh | bashDownloads a Docker Compose file with everything: LLM (Ollama), STT (Parakeet), TTS (Edge TTS), embeddings, Redis, PostgreSQL. Add --gpu for NVIDIA acceleration.
docker run -d --name magec \
-p 8080:8080 -p 8081:8081 \
-v $(pwd)/config.yaml:/app/config.yaml \
-v magec_data:/app/data \
ghcr.io/achetronic/magec:latestCreate backends, agents, and clients from the Admin UI. See the Docker Quick Start guide.
Download from Releases, extract, and run:
./magec --config config.yaml
Ideal for local MCP tools (filesystem, git, shell). See the Binary Installation guide.
Admin UI → http://localhost:8081 · Voice UI → http://localhost:8080
- Multi-agent — Per-agent LLM, memory, voice, and tools. Hot-reload from the Admin UI.
- Agentic Flows — Visual drag-and-drop editor. Sequential, parallel, loop, nested.
- Any backend — OpenAI, Anthropic, Gemini, Ollama, or any OpenAI-compatible API.
- MCP tools — Home Assistant, GitHub, databases, and hundreds more via Model Context Protocol.
- Memory — Session (Redis) + long-term semantic (PostgreSQL/pgvector).
- Voice — Wake word, VAD, STT, TTS. All server-side via ONNX Runtime. Privacy-first.
- Clients — Voice UI (PWA), Admin UI, Telegram, webhooks, cron, REST API, Slack.
See all screenshots in the documentation.
- Multi-agent system with per-agent LLM, memory, and tools
- Visual flow editor (sequential, parallel, loop, nested)
- MCP tool integration (HTTP + stdio transports)
- Voice UI with wake word detection and VAD
- Telegram client with voice support
- Long-term semantic memory (pgvector)
- Session memory (Redis)
- Webhook and cron clients
- Admin UI with hot-reload
- Secrets management (encrypted storage for API keys and sensitive credentials)
- Slack client
- Context window management — automatic summarization when approaching token limits (experimental)
- Discord client
- Expose agents and flows with A2a (Agent-to-agent) protocol
Full docs at magec.dev/docs — installation, configuration, agents, flows, backends, memory, MCP tools, clients, voice system, and API reference.
- Go 1.25+
- Node.js 22+ (for UI builds)
- Docker (for infrastructure services)
| Command | Description |
|---|---|
make build |
Build frontend UIs + embed models + compile server binary |
make dev |
Build all and start server |
make dev-admin |
Start Admin UI dev server (Vite, hot-reload) |
make dev-voice |
Start Voice UI dev server (Vite, hot-reload) |
make swagger |
Regenerate Swagger docs |
make infra |
Start PostgreSQL + Redis |
make ollama |
Start Ollama with qwen3:8b + nomic-embed-text |
make docker-build |
Build Docker image (current arch) |
make docker-buildx |
Build multi-arch image (amd64 + arm64) |
make clean |
Remove build artifacts |
| Dependency | Purpose |
|---|---|
| google.golang.org/adk | Google Agent Development Kit |
| modelcontextprotocol/go-sdk | MCP client |
| yalue/onnxruntime_go | ONNX Runtime for wake word / VAD |
| mymmrac/telego | Telegram bot |
| achetronic/adk-utils-go | ADK providers, session, memory |
| Who | What |
|---|---|
| @travisvn | Built the ARM64 Docker image for OpenAI Edge TTS in record time. This is the local TTS service we recommend — it exposes an OpenAI-compatible API (/v1/audio/speech) that uses Microsoft Edge's free neural voices under the hood, so Magec can use it as a drop-in replacement for OpenAI TTS. |
Apache 2.0 — Alby Hernández
If you find Magec useful, please ⭐ star this repo — it helps a lot.