Give it keyframes. Repair only the frames that break.

FrameLab is not a NLE, not a generator site. It is a timeline of frames with a graph behind them — onion skin, keys, inbetweens, pixel consistency, and an MCP protocol so an agent can work the same commands you do.

Inbetween: Motion plan + linear-blend
Consistency: Pixel metrics, real
Agents: MCP + scopes

Onion skin of a 2-bounce squash-and-stretch. Same shot ships as a sample timeline after you sign in.

What is loaded

Pixel MAE, histogram flicker, 16×16 block matching, motion-plan inbetweens (linear-blend candidate), neighborhood repair, Grok vision on frames you pick.

What is reserved

SAM 2, RTMPose, SEA-RAFT, LocoTrack, Depth Anything, RIFE, Wan. Adapters exist. They return MODEL_NOT_AVAILABLE until a checkpoint is registered.

How agents work

Issue an MCP token in the studio. POST /api/mcp with Bearer. Same commands as the UI — analyze_consistency, generate_inbetweens, repair_frame_range — with scopes and an audit log.