STARTUP.md

Starting AgentCreator

Prerequisites

• Docker Desktop running (open from Start menu, wait for it to fully start)
• Python 3.13 installed (already on PATH)

Step 0: Lock GPU clocks (important for speed)

Run this every time after a reboot — prevents Windows from throttling the GPU:

nvidia-smi -lgc 1500,3090

Without this, inference is ~5x slower.

Step 1: Start vLLM (GPU inference server)

vLLM auto-starts with Docker Desktop (restart: unless-stopped). Verify it's ready:

curl http://localhost:8000/health

If it's not running:

cd C:\vibercoded\gpu-machine
docker compose up -d

First boot takes ~2-3 minutes to load the model into GPU memory.

Step 2: Start AgentCreator

cd C:\vibercoded
python -c "from agent_creator.main import main; main()"

Server starts at http://localhost:7000.

Step 3: Use it

• Dashboard: http://localhost:7000
• Create agent: http://localhost:7000/agents/new (describe in English or paste YAML)
• Topology view: http://localhost:7000/topology
• API docs: http://localhost:7000/docs

Starting/stopping agents

Through the dashboard, or via API:

# Start
curl -X POST http://localhost:7000/api/agents/{name}/start

# Stop
curl -X POST http://localhost:7000/api/agents/{name}/stop

# Build from spec
curl -X POST http://localhost:7000/api/agents/{name}/build

# Health check
curl http://localhost:7000/api/agents/{name}/health

Stopping everything

1. Stop AgentCreator: Ctrl+C in the terminal
2. Stop agent containers: use Docker Desktop or docker stop $(docker ps -q --filter "name=ferite-agent")
3. Stop vLLM: docker stop gpu-machine-vllm-1

Ports

Service	Port
AgentCreator UI + API	7000
vLLM (LLM inference)	8000
Agent containers	9000-9200