Start the local
workbench.
By Alvin
Research Relay lets you watch a question move through several AI roles. Start with Scripted test bench: it uses prepared replies, so your first run needs Python and a browser. You can try an installed local model afterward.
1 / Open Terminal in the extracted pack
Follow page 5 and confirm that ls shows programs and lab. Keep this Terminal window open while the workbench runs. We recommend the walkthrough edition below, which corrects a model-label problem found during these checks. The original program remains in programs/ as a baseline.
python3 lab/beginner-walkthrough-2026-11-15/research-relay/relay.py
2 / Open the printed address
Terminal prints Research Relay: http://127.0.0.1:8802. Open that address in a browser on this same Mac. Here 127.0.0.1 means the Mac, and 8802 identifies the workbench. Ollama uses a separate local port, usually 11434; the app contacts it for you.
3 / Choose the prepared-reply route
Leave Run mode on Scripted test bench. The fictional notes already fill the JSON box: structured text with named fields. Keep them for this run. Click Create run. Expect PLAN, 0 / 5 calls and 0 evidence notes. A role is a job, like planning which notes to use.