AI models forget. Frameloop remembers.
Veo, Seedance, Kling, and their peers generate a clip and move on. Frameloop keeps the record around it — why a shot exists, what it cites, how it cuts for each platform, where it publishes, and what to change on the next pass.
Private controls. Public receipts.
The public site does not expose the local cockpit, shell actions, or project files. It publishes the useful trace: intent, source mix, render path, output variants, and credits.
Every work has structured state: shots, source kinds, durations, prompts, assets, takes, outputs, costs, and review notes.
The loop routes tasks to Seedance, OpenAI image generation, NASA/public-domain sourcing, Manim, Remotion, ffmpeg, and review skills without pretending any one model is the whole product.
One creative direction can become a landscape master, short-form vertical cut, watch page, description templates, proof artifacts, and partner CTA.
Every action is a scriptable skill — the same ones an AI coding agent (Claude Code, Codex, or your own) can run directly against a plain frameloop.json manifest on disk. No hosted API to integrate against: point an agent at a project and it reads and drives the same state the cockpit does.
| Direct generator | Frameloop |
|---|---|
| Prompt in, clip out. | Brief, shot plan, source graph, references, render passes, mux, captions, aspect variants, proof. |
| State lives in chat history and downloads. | State lives in a manifest and cockpit that can be reviewed, rerun, and published from. |
| Platform cuts are manual afterthoughts. | Platform fan-out is part of the release plan. |
| One clip at a time, no receipts. | Batch shot manifests with cost, provenance, and review gates before anything publishes. |
| Closed to automation — a UI or nothing. | Open to your own agent — skills and manifest state a coding agent can drive directly. |