Public marketplace of steering packs. Fork, star, PR. A researcher publishes a 'reduce overconfidence' pack; 500 users apply it with one click to their Qwen3.6 deployment the same day.
Every recipe is a signed bundle: SAE pointer, feature IDs, alpha values, target hook paths. `openinterp apply <url>` patches your transformers model with a monkey-patch hook — no retraining, no fine-tune, no finetune weights to store.
Recipes earn a robustness score based on causal validation across prompts. Red = unstable. Green = reproduces across seeds + datasets. Users see the score before applying.
Recipes that amplify dangerous capabilities (jailbreak patterns, CBRN uplift) are flagged via the Watchtower safety-watchlist features. Flagged recipes are private-by-default and require case-by-case review.
We prioritize researchers, educators, and safety teams who will use it publicly. Tell us what you want to build; we'll reach out when the beta opens.