Point your API calls. Own the model.

$ curl https://kolm.ai/v1/route/chat/completions \
  -H "Authorization: Bearer $KOLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"namespace":"billing-agent","model":"openai:gpt-4.1-mini","messages":[{"role":"user","content":"classify this ticket"}]}'

$ curl https://kolm.ai/v1/capture/export \
  -H "Authorization: Bearer $KOLM_API_KEY" \
  -o billing-agent.jsonl

$ curl https://kolm.ai/v1/compile \
  -H "Authorization: Bearer $KOLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"task":"billing-agent","examples_file":"billing-agent.jsonl","target":"local-cpu"}'

Here is what each of those three calls does, the response you should see, and roughly how long it takes. A namespace (a named bucket for one agent's traffic)A namespace is a label you attach to one agent's calls, for example billing-agent. Captures, evals, and the compiled model are all scoped to it, so two agents never mix. keeps each agent's traffic, examples, and compiled model separate.

{ "id": "chatcmpl-...", "choices": [ { "message": { "role": "assistant", "content": "refund_request" } } ], "x_kolm": { "captured": true, "namespace": "billing-agent" } }

{"namespace":"billing-agent","input":"classify this ticket","output":"refund_request","ts":"2026-06-14T10:02:11Z"}
{"namespace":"billing-agent","input":"where is my order","output":"order_status","ts":"2026-06-14T10:02:14Z"}

{ "id": "cmp_018x", "status": "queued", "target": "local-cpu", "artifact_url": "/v1/compile/cmp_018x/.kolm" }

Want the bigger picture before you build? See the whole compiler on /platform, the exact .kolm format and its Ed25519 signature on /spec, where artifacts run on /runtimes, and check any signed model offline at /verify.

• Set up your accounts and connections: workspace, project, environment, source, connector, credential, and provider.
• Decide what gets kept and how it is tested: capture policy, redaction policy, retention policy, eval suite, failure taxonomy, and regression set.
• Ship the result: compile run, artifact, manifest, target runtime, receipt, export, and release checks.

• POST /v1/route/chat/completions - point your existing OpenAI-compatible traffic at Kolm, no SDK rewrite.
• POST /v1/gateway/dispatch - send calls to your providers with capture, caching, fallback, and quality routing built in.
• POST /v1/capture/log, GET /v1/captures/list, GET /v1/capture/export, POST /v1/capture/bulk - pull the behavior you captured into review and training.
• POST /v1/compile/estimate, POST /v1/compile/preview, POST /v1/compile, GET /v1/compile/:id/.kolm - turn that behavior into a signed model you own.
• GET /v1/devices/detect, POST /v1/devices/recommend, POST /v1/devices/:id/install, POST /v1/serve - run your model on the hardware you already have.

• Every .kolm artifact is Ed25519-signed with content-addressed receipts - signed so anyone can verify it.
• Export the eval reports, manifests, and receipts and attach them to your release as proof.
• Verify any artifact at /verify, and see every check we run at /checks.

$ kolm signup --email you@company.com
$ kolm login --key ks_...
$ kolm compile "billing-agent" --target local-cpu
$ kolm verify artifact.kolm
$ kolm serve artifact.kolm