Best-Effort Clipping, Face Tracking and Fixes

This release changes what happens when a video doesn't fit the clipping brief you gave it instead of coming back empty, it now returns the best clips it can and tells you what it adjusted. It also brings face tracking to the editor as a per-segment control, adds a Connect your AI setup flow with a dedicated API Keys page, and fixes a set of editor saving and caption issues.

Clip Generation

Clip generation is now best-effort. Videos that used to struggle outright now deliver clips.

What changed

Impossible briefs are reconciled instead of rejected. Asking a 30-second video for two 30–60 second clips used to produce zero clips and an invalid project. Requested clip lengths and clip counts are now checked against the video's actual runtime before generation starts, and adjusted to something the video can support.
Short videos no longer come back empty. When no requested duration window fits the runtime, one that does is substituted, and the clip count is capped to how many non-overlapping clips the video can actually hold.
Topic field keywords now match more reliably (even with extra spacing), and requests for "one-liners" are automatically recognized as quote-style clips.

Face Tracking

Face tracking (center stage) is now a per-segment setting in the editor rather than all-or-nothing for the clip.

Per-segment toggle. Each full-frame speaker segment has its own Face tracking switch in the video settings panel, so you can let the frame follow the speaker on some segments and leave others fixed.
Availability is based on the footage. Segments where a face isn't visible for most of the duration show the toggle as unavailable, with a tooltip explaining why, instead of enabling a pan that has nothing to follow.
Bulk toggle in AI Tools turns tracking on or off for every eligible segment at once.
Fine-tune controls. With tracking on, a Fine-tune tracking panel exposes follow speed and dead zone, with a Reset back to defaults.
Repositioning asks first. Manually moving or resizing a tracked segment turns tracking off, so it now confirms before doing that. Turning tracking on tells you up front that it restores the clip's original size and position.

Connect your AI

API Keys now has its own sidebar entry, instead of being reachable only from inside settings.
The Connect your AI dialog covers ChatGPT, Claude, Cursor, and OpenClaw, each with the setup path that fits it: copy the server URL and open the client, copy the CLI command, or a one-click Add to Cursor install.

Editor Fixes

Saving a clip now confirms it worked. Saves always succeeded but the editor never said so. You now get a success toast, and failures including asset upload errors, which previously passed silently, are surfaced.
Drafts saved from later pages of the clip list no longer look unsaved. Only the first page was being refreshed, so a saved draft further down reverted to its old state until a hard refresh. The list also refreshes when you return to the tab, for the duplicated-tab workflow.
Fixed the editor hanging at "Loading 0%" when opened in a background tab. Chrome won't load video for a hidden page, and the editor waited on it forever; it now recovers when the tab becomes visible, or proceeds without the thumbnail.
Caption font size no longer resets when you clear the field. Clearing the input and clicking away saved an empty value; it now restores the previous size. Zero and negative sizes are rejected, and clips saved with a bad size render at a readable minimum instead of disappearing.
Fixed a crash when adding a segment after the transcript changed underneath the dialog.
Fixed the word menu appearing for a word that no longer exists in the transcript.
Fixed a rendering failure on very short overlays. A template overlay or asset shorter than a single frame on a low-frame-rate source was silently dropped, or failed the export outright; these are now kept at one frame.

Other Fixes

Malformed video links are caught when you paste them. A link with junk glued to the front, the kind produced by a bad copy-paste used to pass validation, and a mangled YouTube ID would be truncated until it matched some video, so you'd get details for a video you didn't ask for and a failure much later with no useful explanation. Links are now cleaned and validated up front, with a clear message when one can't be used. Bare domains without https:// are accepted, tracking parameters are stripped, and playlist links are rejected on the API as they already were in the app.
Fixed long, high-resolution uploads getting stuck. A 4K upload over about two hours could exceed the fixed processing time limit mid-encode and stall permanently, leaving the project processing and every clip in it unfinished. The limit now scales with the length and resolution of what's actually being processed.