Agent Lab

Agent Lab is the default way to use AITracer: run agents on Ollama, save every run as a trace, and export data for training.

Prerequisites

1. Install Ollama on the machine that will run models.
2. Pull a model:

ollama pull llama3.2
ollama serve

3. Sign in to AITracer and open Dashboard → Run agent (/dashboard/lab).

If the app runs in Docker/Podman, set OLLAMA_BASE_URL so the container can reach the host, for example:

OLLAMA_BASE_URL=http://host.containers.internal:11434

Run & trace

In Run agent:

1. Pick a local model (loaded from /api/local/models).
2. Set an action name (used to filter training exports later).
3. Enter a prompt and click Run & trace.

Each run calls POST /api/local/run with recordTrace: true. Traces are stored with costUsd: 0.

Agent chat

Dashboard → Agent chat (/dashboard/ai/chat) runs multi-turn conversations through the same local backend. Every assistant turn is traced as agent_chat.

API (local runs)

curl -X POST http://localhost:3000/api/local/run \
  -H "Content-Type: application/json" \
  -H "Cookie: <session>" \
  -d '{
    "prompt": "Explain recursion in one paragraph.",
    "model": "llama3.2",
    "actionName": "agent_lab_run",
    "recordTrace": true
  }'

List models:

curl http://localhost:3000/api/local/models -H "Cookie: <session>"

Next steps

• Coach — optional paid review of local drafts
• Training data — export JSONL or build an Ollama model
• First trace — trace lifecycle and API ingestion