The page is the tool.

Open this page or the studio in ChatGPT desktop on GPT-5.6 Sol or Terra. Site tools land on the still you can pin. Luna has WebMCP off. Codex cannot see the photograph — `see_still` tells it the look. When it writes, the picture on this page changes.

Judge path

Open slate-webmcp.vercel.app or /studio in the in-app browser. The still on the page is live.
Ask what is on screen. It should describe the laugh — a woman mid-laugh, 05-A, pinned — and say what to do next.
Ask for a recut suggestion. It should leave the laugh and propose the hand and the landfill.
Apply that recut. Rehearse the laugh. The still should refuse on 05-A. The landfill should still trim.
Clap on the page. Export waits for a human.
Watch from the top. Same stills. New caption. Same pin.

Say this

Tools

help

The judge path and the next move. Call this if you are lost.

see_still

What the photograph shows. Also Codex's last beat.

look

Same as see_still.

suggest_cut

Propose the next recut. Does not write.

apply_suggestion

Apply that recut. Rehearse also tries the laugh.

mark_cut

Arm the clap. Same as request_export.

clap

Commit after arm. Human can clap too.

hold

Same as trim_shot. Fails on a pin.

Same as lock_shot.

replay

Play the cut from the top.

direct

Several named edits. Pins refuse a line and continue.

get_project

Read the whole cut.

find_shot

Look up a still. Do not move.

select_shot

Land on the same still.

set_caption

One call. Pass query. Fails on a pin.

trim_shot

Seconds, ms, or delta. Fails on a pin.

split_shot

At the playhead, or mid if you name another.

move_shot

Reorder. Pass after, not just an index.

get_edl

Read the cut. Does not clap.

what_changed

Diff the last edit.

lock_shot

Pin it. Writes then refuse. Pins may still move.

Every tool registers in JavaScript on the live page. ChatGPT does not read HTML forms. Same reducer. No second backend. On the page: space plays, W watches from the top, arrows step, L pins, M marks, C claps.

Open the laugh