What is the Aigentron Dev Server?

It is a self-hosted orchestrator. A Docker Compose stack starts several coding agents, gives each one an isolated workspace and coordinates their task queues, so a fleet of agents can work on a repository instead of a single chat session.

Do I need an API key to run it?

Yes. The agents call a model provider, so you supply your own Anthropic or OpenAI key in the local .env file. The key stays on your machine and is never uploaded to arttron.dev.

What are the prerequisites?

Docker Desktop, or Docker Engine with Compose v2 on Linux, plus roughly 8 GB of free disk for images and workspaces, and outbound HTTPS access so the agents can reach the model API.

Aigentron Dev Server - Self-Hosted AI Coding Agent Server

Version

Local installation

Install and start the stack with the one-line installer, or clone the source and run it yourself.

1. Quick install with install.sh

The installer downloads the tagged release, builds the image locally and starts the stack; the profile/image is chosen by the installer. Re-running upgrades in place. Pin an explicit version with VERSION=0.1.0 sh install.sh.

curl -fsSL https://raw.githubusercontent.com/Arttron/aigentron/main/install.sh | sh

Or clone the source (optional)

git clone https://github.com/Arttron/aigentron.git
cd aigentron

Prerequisites

Docker Desktop, or Docker Engine + Compose v2 (required)
Optional: Ollama running on the host (not in a container) for the 'routine' model tier — just one provider option, not a requirement. The orchestrator reaches it at http://host.docker.internal:11434
ANTHROPIC_API_KEY — required for the 'complex' model tier (cloud Claude)
Optional: a tool-capable model in your host Ollama (default qwen3-coder:30b, set via ROUTINE_MODEL in .env) — only if you choose Ollama
Optional: a one-time Docker network so agents can reach project databases by service name (docker network create lds-agents) — only for the compose/agent workflow
If you do use Ollama, avoid qwen2.5-coder — it does not emit structured tool calls and breaks the agent SDK. qwen3-coder:30b is the supported default.

2. Optional: local models with Ollama

Only if you use Ollama as your provider. Pull into your host Ollama (not into a container); set ROUTINE_MODEL in .env if you pick a different model.

ollama pull qwen3-coder:30b

3. Optional: external Docker network

One-time setup, only needed for the compose/agent workflow so agents reach project databases by service name.

docker network create lds-agents

4. From source: build and start

For the source checkout: equivalent to docker compose up -d --build. Brings up Postgres, Redis, LiteLLM (Anthropic-compatible proxy in front of Ollama), the orchestrator (:3001), and the dashboard (:3000), all bind-mounted for hot-reload. First boot is slower — the orchestrator compiles packages/shared and packages/agent-runner before its healthcheck passes.

make up

5. Open the dashboard

Dashboard → http://localhost:3000 • Orchestrator API → http://localhost:3001 The defaults are shown above; the dashboard port can differ in some setups — override it via the port mapping (compose or docker run -p). Optional: docker compose --profile mcp up -d for Playwright MCP and a dev server for frontend/design agents.

Useful: make logs | make ps | make migrate | make down | make clean (⚠️ wipes DB data). More detailed setup, model routing, and approval gate docs are in the repo.

Back to home