Point your provider client at kolm. Every call is captured, hashed, and dropped into this inbox. Keep what matters, discard what's noise, promote a winner into a recipe. When a namespace hits 1,000 pairs, distill it into a signed .kolm Specialist that runs on your own model — no upstream tokens, no per-call latency, full receipt chain.
The captures inbox is gated by your API key. Sign in or get an API key and your live captures land here.
| Step | What happens |
|---|---|
| 1. Capture | Your client hits kolm.ai/v1/capture/{openai,anthropic}. We forward upstream, return the response, and record (model, prompt_hash, output_hash, latency, ts). |
| 2. Triage | Keep — lock into the next distill batch. Discard — remove from training. Promote — deep-link into the recipe playground to lift this single observation into a reusable spec. |
| 3. Distill | At 1,000 kept pairs in a namespace, the namespace flips to ready. One click fires POST /v1/specialists/auto-distill: the trainer bridge fits a LoRA, stamps a K-score, and returns a signed .kolm artifact. |
| 4. Switch | Re-point the client at your own kolm serve. Per-call cost drops to electricity; the artifact ships with the receipt chain that proves how it was built. |
Your upstream provider key (forwarded in x-upstream-api-key, dropped after the call). Cross-tenant data — the inbox is scoped to your tenant id by your API key. Raw bodies older than your retention window: hashes only, configurable per-namespace.