What changed in FramePilot.

New

• The site opens with a ripple delete. The editors you already know appear on a track, a playhead sweeps through, and every one of them lands in the bin in the corner. FramePilot is the clip that stays and becomes the logo. It plays once, you can skip it by scrolling, and it stays still if your system asks for reduced motion.
• Every page is on the same timeline. Pricing is one clip with one price, downloads are three tracks with your platform highlighted, and the docs, blog, and changelog sit on the same ruler. The bin stays in the corner of the home page, with the retired tools still in it.

Fixed

• A full-song text block can no longer pass as finished lyrics. Caption checks now require separate readable cues with trustworthy timing, and title cards no longer inflate the reported caption count.

Improved

• Visual quality claims now require visual review. If a preview cannot be rendered, the agent clearly leaves lighting, colour, caption readability, typography and motion marked as unchecked instead of saying every check passed. Caption styling also starts restrained instead of piling multiple animations, boxes, outlines and accent colours onto the same lyric treatment.

Fixed

• Long AI edits use far less memory. Large undo histories no longer get copied into every project read, so complex edits can keep running without exhausting the desktop app.
• Reloads no longer cause repeated delivery errors. If the editor reloads while an AI edit is in progress, FramePilot cleanly detaches the old view while the edit remains available to reconnect.

Improved

• Long footage opens and previews with far less work. FramePilot now streams full-length originals as you edit, while automatically using lightweight preview media when it is available.

• Large timelines stay responsive near the playhead. Thousands of cuts, captions, and effects elsewhere in the movie no longer add work to every displayed frame.

• Long caption tracks stay quick to edit. Only the cues currently on screen stay active, so a full movie's subtitles do not slow down typing, timing, or scrolling.

• Caption adjustments stay fluid. Dragging size, width, or line-height controls previews immediately and becomes one undo step when released, instead of creating an edit for every pointer movement.

• The AI panel remains responsive during long jobs. Multi-step edits keep their place and continue showing progress without repeatedly rebuilding the editor around them.

Fixed

• Paused means idle again. The timeline and effect preview stop running display-rate update loops when playback is paused, reducing CPU use and fan noise between edits.

• Agent steps no longer reset the workspace. Applied steps update the open project in place, while rapidly arriving text is grouped into smooth updates so controls remain usable.

• Long AI sessions no longer run out of application memory. Large undo histories stay in the editor instead of being copied into every AI request, and growing agent activity is recorded without repeatedly reloading the entire session.

• Old undo data no longer stalls saves or agent steps. The current session keeps its full undo stack, while the project file retains a bounded recent undo window so autosave stays responsive on large movies.

Improved

• Clean, familiar selection controls. Objects and the background frame now use slim white borders and compact white handles that stay readable over bright or dark footage.

• Deliberate object picking. A single click selects the background; double-click the text you want to work on to isolate it. Keyboard users can select the focused object with Enter or Space.

Fixed

• Timeline selections now appear on the video. Select a text object in the timeline and its editable bounds immediately appear in the program monitor, ready to move, resize, or edit.

New

• AI emphasis finds the words that carry the idea. Your selected AI model weighs meaning, emotion, pauses, delivery and sentence structure, then gives a small number of anchor words the visual hierarchy they deserve. You can replace every suggestion, and captions still work with a clearly labelled local fallback when the AI is unavailable.

• Ask the editing agent to design captions end to end. It can now choose from the real font and template library, emphasize only words that are actually spoken, and control placement, size, rotation, width, alignment, spacing, backgrounds and animation. Every change remains reviewable, editable and undoable just like a manual caption adjustment.

• Twenty-two creative fonts, ready everywhere. Pick from familiar sans, bold display, editorial serif, monospace and handwritten families. They are included with FramePilot, so the same typography appears in preview and export on another machine.

• Captions are editable objects in the Program monitor. Drag one into place, resize its wrap width, rotate it, or double-click to rewrite it. Alignment, line height and safe-area behavior carry through to export.

• Six new production looks. Semantic Anchor, Editorial Contrast, Compact Tier, Kinetic Stack, Handwritten Zone and Social Headline bring creator, podcast, editorial, lyric and short-form personalities, not just different fonts.

Improved

• Segmentation now composes around emphasized words. Cue and line breaks still follow pauses, punctuation and reading speed, while reserving enough visual room for the anchor word so phrases stay balanced and readable.

• Generated captions remain a starting point, not a locked result. Rewrite, split, merge, restyle, reposition and undo without returning to transcription.

New

• TwelveLabs is now a transcription provider. Choose it in Settings and reuse the same key as Media intelligence. FramePilot indexes only the clip you chose, then brings its native word timing back into the editor.

Improved

• Local transcription uses a substantially stronger multilingual model. Word timing follows the spoken audio, including silence before the first word, so generated captions land where the voice actually begins.

• Caption styles are quicker to compare. Browse the complete catalog with All, narrow it by category, or search by a style's name or behavior. Richer cards preview the real caption treatment when you hover or focus them.

• Waiting feels dependable. Fast cached transcripts do not flash a loader; longer jobs show honest provider-specific status and elapsed time without a fake percentage.

Fixed

• A TwelveLabs key no longer changes providers behind your back. The provider you choose is the provider FramePilot uses. If indexing or transcription fails, the existing transcript stays intact and the editor explains what happened.

Improved

• Effects now appear in the preview you use by default. Looks and stylised effects update directly in the program monitor, so you can judge the edit before exporting it.

• Every monitor action stays within reach. Orientation, before-and-after compare, loop, composition grid, safe areas, fit and zoom, fullscreen, frame stepping, seeking, and timecode stay available throughout editing.

• One dependable preview, with no engine switch to manage. The older preview option and its Settings control are gone, so effects and compositing always use the same monitor.

Fixed

• Portrait and custom canvases keep their true shape. The picture now fits cleanly inside both the available width and height, without stretching or losing its orientation when panels move or the project switches between landscape and portrait.

• The progress bar now sits where it belongs. It spans the monitor above the centered playback controls instead of being squeezed against the right edge.

• A broken source no longer snaps your edit back to the start. Its error stays visible while timeline timing, captions, effects, and history remain usable.

Improved

• The agent keeps going when a plan still has work left. If a long edit produces one safe batch and then tries to stop early, FramePilot gives it a focused opportunity to finish the next pending step. "All checks passed" now means the requested result was actually reached, including an explicit duration such as a 30-second video, while any safe partial changes remain available for review.

• Live activity is easier to follow. Thinking, tool use, editing, and verification now animate beside the message box where you can stop or steer the run. Earlier thoughts stay settled as "Thought for Ns" instead of appearing to start thinking again when a new step begins.

Fixed

• Preview playback stays smooth across mixed footage. Video-only stock clips, clips with sound, still images, gaps, and a music bed now share one steady timeline clock, so the program monitor no longer flashes black and the playhead no longer jumps back and forth. High-refresh displays also reuse each real video frame cleanly, while the playhead stays crisp on the screen's physical pixel grid.

• Opening Inspector or Transcript no longer interrupts AI chat. The agent keeps working in the background, and returning to AI restores the same active conversation and draft.

• The context figure stays steady for your message. Internal planning and tool requests no longer replace the number with their smaller or larger prompt sizes, so ordinary agent work does not look like conversation memory is disappearing.

• GLM-5V-Turbo uses its full supported capacity. FramePilot now recognizes both the configured glm-5v-turbo name and its provider-qualified id, with a 200K-token context window and up to 131,072 output tokens.

Improved

• Your active plan stays within reach without taking over the sidebar. Every run starts with a compact, collapsed plan below the AI header, showing only the latest task and progress. Open it whenever you want the full checklist.

Fixed

• A late polish step no longer throws away a finished montage. Transitions, grading, and motion accents now see the clips created earlier in the same run and are included in one final check. If that optional refinement still cannot produce a safe result, FramePilot clearly marks the warning and keeps the already validated cut for review instead of terminating the whole edit.

• Montage clips no longer fail because a timeline track was mistaken for a media file. FramePilot now keeps those identities separate, checks the whole proposed edit before moving on, and gives the agent the exact problem to correct. If the edit still cannot be made safely, assembly and verification do not start and the failed step explains why.

• A montage no longer appears complete when no edits were produced. FramePilot gives an empty proposal focused, bounded correction attempts. If it still cannot produce the requested timeline changes, the exact step fails clearly instead of checking off assembly and verification.

• Complex planned edits no longer stop just before the proposal is assembled. The steps already connected in a plan now carry their results forward automatically, so a montage can move from footage analysis through assembly without failing over a missing internal reference.

• An AI run can no longer keep editing after losing its objective or committed plan. FramePilot now saves the request and plan before applying changes, links each operation back to that plan, and avoids repeating an operation during recovery. If the durable run state is missing, stale, or inconsistent, later edits pause safely and the existing timeline is preserved for review.

• "All checks passed" now requires real completion evidence. Verification reconciles the committed deliverables with the edits that actually landed. An incomplete or untraceable run ends honestly instead of presenting a successful summary.

New

• Know when a request is getting crowded. A small ring beside Send shows how much of the current AI request's context is in use. Hover or focus it for the used and total token counts plus the exact amount remaining.

Fixed

• Long AI edits keep moving. If the fast planning path cannot execute a valid graph, FramePilot continues the same request with its general editing agent instead of stopping. The selected provider stays consistent throughout, oversized timelines keep coverage from beginning to end, and incompatible Ollama sampling settings recover automatically.

Fixed

• No more surprise empty chats mid-edit. When the agent applies changes to your timeline, the conversation now stays visible with its live progress and working Stop control instead of suddenly jumping to a blank new chat.

Improved

• Preview controls stay out of your way. Playback remains centered below the picture, while canvas, guide, zoom, and fullscreen controls sit together at the top right.
• Panels adapt as you resize them. Transcript phrases wrap naturally, and the Assets header, filters, and import controls reorganize cleanly in a narrow rail.

Fixed

• AI chats stay with their project. Switching projects no longer brings unrelated conversation history along with it.

Improved

• Auto mode is now trustworthy. When you authorize automatic editing, each change is saved through the same protected project workflow and appears in your workspace immediately.

• Undo a whole AI run in one step. Multi-stage edits from one request are grouped together, so a single Undo returns you to the state before the run.

Fixed

• Newer work is never silently overwritten. If your project changes while an AI edit is in progress, FramePilot either applies the edit safely to the latest version or asks for a fresh plan.

• AI state survives reopening the app. Project versions and completed edit decisions remain consistent after a restart, preventing the workspace from showing stale execution results.

New

• Show 2 alternatives (Edit mode). When you ask the AI for a quick edit, a new "Show 2 alternatives" toggle lets it propose two independent takes instead of one. Flip between "Take A" and "Take B" in the review card (same before/after player, re-pointed at whichever take you're looking at) and accept the one you like. The other is discarded, never left half-applied.
• Honest about cost. A second take is a second real AI call, so it's off by default and the toggle says so up front. When you turn it on, the usage chip reflects the combined cost of both takes, not just one.
• Only where it makes sense. Deterministic one-click actions (remove silences, add captions, fix pacing, and the rest) never show this toggle. Running the exact same recipe twice would just give you the identical result, so there's nothing to compare.

New

• Quick command palette (⌘K / Ctrl+K). Press ⌘K anywhere in the editor to open a command palette. Select a clip first and type what you want changed, like "tighten this up" or "make the cut punchier", and it's sent straight to the AI as an edit scoped to that clip, no need to open a chat first.
• "Ask AI about this clip." Right-click any clip on the timeline for a new "Ask AI about this clip" option. It selects the clip and opens the same quick command palette, already scoped to that moment, click, describe the problem, and let the AI adjust it.
• Keep refining in place. After a quick edit lands, the AI sidebar opens with the same conversation ready to go, so you can keep asking for tweaks without starting over.

New

• Cross-project memory. A new Settings → Memory panel lets you set your editing defaults once, who your videos are for, your brand and caption style, how tight you like the cut, and where you export, and every new project inherits them. A project's own choices always win, so nothing is ever overridden without your say-so. Your defaults follow you, not the file.
• Save a run as a recipe. Save an edit you like as a named workflow and it becomes a one-word shortcut: type its name later and it replays instantly, with no AI call at all. The more you teach FramePilot, the faster and cheaper your everyday edits get. Manage your saved workflows in Settings → Memory.

Improved

• Sturdier agent runs. When something goes wrong mid-run, a slow model, a step that can't complete, the agent now recovers on its own: it retries, falls back to an instant recipe or a cheaper model, works around the failed step, or pauses honestly for your review. It never pretends a step succeeded.
• Instant commands stay instant, even as FramePilot grows. Common edits like "remove the silences", "add captions", and "export for Reels" run with no AI call, and the groundwork is now in place to keep them fast and cheap no matter how many new capabilities we add.

New

• Add text by dragging. Grab Text from the Overlays panel and drop it anywhere on the timeline to create an overlay right where you want it.
• Edit right on the video. Select a text overlay and work with it directly in the preview, drag to move it, pull a side handle to set how wide it wraps, and double-click to type. No fiddly forms.
• Style it your way. A new Text panel puts everything in one place: content, colour, font and weight, size, alignment, a background, on-screen position, and in/out animations with timing. The preview shows exactly how it will look and moves, even as you scrub.
• Pick your transitions. Clicking the + between two clips now opens a quick menu of every transition (fade, cross-dissolve, push, zoom, blur) instead of just dropping in a default.
• A composition grid. Turn on a rule-of-thirds grid in the monitor to line up your shots.

Improved

• A calmer preview. The area around your video now matches the app instead of harsh black, the frame fits perfectly at any window size, and fast cut-heavy edits play back more smoothly.
• Status where you expect it. "Saved to…" and other messages now appear quietly in the header instead of popping up as a separate toast.
• Snappier zooming. Zooming the timeline in and out repeatedly no longer bogs the app down, thumbnails and waveforms are cached smartly and reused.
• Leaner AI. Behind the scenes the assistant now sends less redundant data with each step, so it works within tighter model limits.

Fixed

• No more empty "proposed edit." When the assistant only made a plan, replied in chat, or hit a snag, it no longer shows an empty edit card with nothing to apply, and the "Apply all" count stays accurate.

New

• Beat detection. Ask the agent for a montage cut to the music and it can now find the beats, it analyzes your audio and gets back every beat plus the song's tempo, so cuts land on the rhythm instead of near it.
• A live to-do list. Ask the agent to plan first and its plan appears immediately as a checklist, you watch each step light up while it runs and check off as it finishes, so you always know where the run stands.
• A closing report. When a run finishes, the agent sums up what it changed, what it skipped, and why, no more scrolling back through the activity to piece together what happened.

Improved

• The agent talks you through the work. Its explanations now appear as chat messages right between the actions they describe. "trimming the silent gaps now…" followed by the trims, instead of being tucked into a side panel.
• Honest activity cards. Every analysis step shows a live timer while it actually runs, tells you exactly what it was asked to do, and shows how long the real thinking took.

Fixed

• No more phantom checkmarks. Analysis steps used to show as done instantly, even when nothing had run. They now run for real, and pressing Stop marks the interrupted step as stopped, never as completed.
• Zooming with thumbnails on is smooth again. The timeline filmstrip no longer re-decodes its images on every zoom tick, so pinch-zooming a big project stays fluid instead of freezing the app.
• Tidier collapsed panels. A collapsed side panel is now a slim, quiet edge that glides open and closed, without a leftover drag handle beside it.

New

• Resize and reposition clips right in the preview. Click a clip in the monitor to get a selection box with corner handles, drag to move, pull a corner to scale. Every gesture is a single undoable edit.
• Canvas presets. Switch your project between 16:9, 9:16, 1:1, 4:5, and 21:9 from the monitor, the preview, guides, and export all follow instantly.
• GitHub Models and GitHub Copilot. Bring your GitHub token and run the editing agent on GitHub-hosted models, straight from Settings → AI.

Improved

• Previews now play lightweight proxies. Imported footage gets a fast preview copy automatically, so scrubbing and cuts stay smooth even with heavy camera files, exports still use your originals at full quality.
• The assistant streams its thinking instantly. Reasoning appears token by token with proper formatting, no more buffered paragraphs or shimmer placeholders, and long sessions stay fast from the first message to the thousandth.
• A quieter, quicker workspace. The AI panel opens first, tool details fold away until you want them, the app launches straight into one composed frame, and zooming a long timeline stays smooth with thumbnails that never vanish.

Fixed

• Stopping the assistant now always stops it, even if you switched models mid-run.
• Cutting from video to a still image no longer flashes, and gaps in the timeline show the empty state instead of a stale frame.

Fixed

• Click a text overlay to edit it. Selecting a text overlay right on the preview now works as you'd expect, click to pick it, drag to move it, pull a handle to resize it, and double-click to change the words. No more clicks slipping through to the clip behind it.
• The shape menu opens where you expect. Choosing your canvas shape, Landscape, Reels, Square and the rest, now opens fully in view instead of tucking behind the video or running off the edge of the window, and it lines up neatly with the playback buttons.
• Tooltips stay on screen. Longer tooltips no longer spill off the side of the window, they wrap and stay fully visible.

New

• Choose how the AI's edits reach your timeline. A new Apply mode control in the assistant sidebar puts you in charge. Keep Manual (the default) and every edit shows up as a card you review and Accept or Reject, just like today. Switch to Auto and each edit lands on your timeline the moment the assistant proposes it, so a multi-step request can build and adjust your video hands-free while you watch. Either way, every edit is checked before it's applied, and a single Undo takes it back.

Fixed

• No more "busy but nothing happened" runs. When you ask the assistant for something, it now either shows you the edits you can apply or tells you clearly that nothing could be applied and why, instead of looking busy and leaving you with no result.
• Long error messages stay tidy. Lengthy messages no longer spill outside their cards in the sidebar, so the panel stays readable.

Improved

• Watch the assistant think. The AI panel now streams the assistant's real reasoning as it works, so you can follow what it's doing instead of staring at generic placeholder text.
• Calmer, more honest progress. We removed the misleading percentages and "100%" markers. AI work doesn't really have a percentage. Now you see a subtle loading indicator while the assistant works, and the panel simply goes quiet when it's done.
• Know exactly what each step did. Every action shows an at-a-glance status, working, done, needs attention, or failed, with a short summary and a tooltip. Open "View details" to see the full picture, copy anything with one click, and if a step fails, hover to see why.
• A smoother place to type. The message box now grows to fit longer, multi-line messages while the add and send buttons stay right where you expect them, plus a clearer stop button for when you want to halt the assistant mid-thought.

Fixed

• AI edits actually go through now. The assistant sometimes rejected its own changes, like adding or trimming a clip, and finished with "no changes" even though it had real work to show. Those edits now apply reliably, so asking the AI to build or adjust your timeline produces an edit you can review.

New

• Find and remove silence, automatically. Ask the agent to clean up your footage and it now pinpoints every silent gap for you, the long pauses, the dead air between takes, the "um, let me start over" moments, so you can trim them in one pass instead of hunting frame by frame.
• Spot every scene change. FramePilot can now detect where each new shot begins, making it effortless to split a long recording into clips, drop in b-roll at the right beats, or jump straight to the moment you care about.

Improved

• Faster, steadier imports. Large clips and long recordings now load more reliably, and FramePilot no longer gets stuck waiting on a file that's slow to read, you get a clear result instead of a spinner that never ends.

New

• FramePilot is officially available. Download the app for macOS, Windows, and Linux, activate it with your subscription, and hand the tedious parts of editing, cuts, captions, silence, pacing, off to an agent that works on a real, professional timeline.
• One simple subscription. Get the whole editor for one price, the AI agent, the render engine, and every new tool we ship while you're a member. No ads, no watermarks, and no per-export fees. Choose monthly or save with annual billing, and cancel anytime.
• Your work stays yours. Editing happens locally on your machine, every change the agent makes is reversible, and every export is checked before it finishes, so you're always in control of the final cut.

Improved

• Effortless updates. New versions install automatically in the background, so you're always on the latest FramePilot without lifting a finger.

Improved

• The agent recovers on its own. If the AI service is briefly busy or your connection stumbles, FramePilot quietly retries instead of giving up, so a momentary blip no longer interrupts your edit.
• It remembers the whole conversation. Ask a follow-up like "now make it vertical" and the agent knows exactly what "it" is. Your back-and-forth builds on itself instead of starting over each time.

Fixed

• No more endless spinners. If a response ever stalls, FramePilot now tells you clearly and offers a one-click retry, rather than leaving you waiting.