Raccord Check for developers

Raccord Check answers one question about a take of AI-generated video, from any model: is this take good? You send the clip and photos of the characters who should be in it. You get back pass, fail or not sure, with every problem in plain words, the time it happens, the frames that show it, a suggested fix and a confidence.

It also saves renders. It checks a start frame before you pay to animate it and a prompt before you render it. It ranks a batch of takes, stops asking for takes once one passes, and stops a drifting render partway on models you run yourself.

Status. Raccord Check is in development. The API opens at launch, expected Q4 2026. These docs describe the working build. To try it early, write to hello@kolamlabs.ai.

Three ways in

curl "https://api.kolamlabs.ai/v1/checks?wait=60" \
  -H "Authorization: Bearer rk_test_..." \
  -F clip=@take.mp4 -F "reference[Maya]=@maya.jpg"

from raccord_sdk import Raccord

verdict = Raccord("rk_test_...").check("take.mp4", references={"Maya": ["maya.jpg"]})
print(verdict.status, verdict.headline)

For AI agents (Claude, Cursor and any MCP client), add the MCP server and ask: “Check take.mp4 against Maya's photos.” See Integrations.

Where to go next

• Quickstart: keys, test mode and a first check in two minutes.
• Reading a verdict: every field, and what each check catches.
• Endpoints: checks, rankings, start frames, pre-flights, extensions and previews.
• Saving renders: the gates, and how to wire them into a pipeline.
• Reference photos: what makes a good reference.
• Webhooks, Errors and retries, Signed verdicts, Data and privacy.
• Accuracy: how well it works, how that was measured, and its known limits.
• Pricing.

Machine-readable: /v1/openapi.json and /llms.txt.