Droid
Configure Factory Droid with a Tokamak custom model and preserve your other models and settings.
Install and launch
Complete Install and connect, using tokamak-stag on staging. Tokamak installs the official droid npm package if missing; existing standalone installations also work. Factory account requirements still apply. Select a Chat Completions model.
tokamak launch droid --model YOUR_MODEL_ID --check
tokamak launch droid --model YOUR_MODEL_IDFor noninteractive execution, place Droid's command after the separator:
tokamak launch droid --model YOUR_MODEL_ID -- exec "explain this repository"
tokamak launch droid -- --helpPersistent configuration
Tokamak merges a custom model into ~/.factory/settings.json and selects it as the default. Other custom models, permissions and reasoning settings remain. This default persists after exit; launch through Tokamak to supply its API key.
The key is passed through the environment and referenced from the model configuration. Droid 0.213 can expand and save that key during startup. Tokamak restores its environment reference after normal exit or spawn failure when the exact key still matches. A forced kill or detached child can leave the key persisted, so treat settings and backups as credentials.
The first original file is saved as settings.json.tokamak-backup. With Droid closed, reconcile or restore that backup manually if you want to undo setup. Avoid concurrent settings edits. Invalid configuration, duplicate managed models or conflicting entries are rejected before writing.
Models, verification and troubleshooting
Use tokamak launch droid --model to change the selected model. Explicit native model choices, worker/spec models and resumed sessions may select other routes; check the active model in Droid. Managed and project policy retain their precedence.
The recorded Windows check validated repeated configuration, version and exec help for Droid 0.213.0. Execution stopped at Factory's missing-account-credentials gate; it did not demonstrate successful inference. Sign in as required by Factory before checking routed requests.
Server inference is metered by Tokamak. No client telemetry exporter is configured by this adapter. Tools, reasoning, images and caching depend on the selected model. See troubleshooting.