Frapea

Places media segments on a timeline. Each clip: mediaRef (from get_media, 'timeline:<id>' to NEST another timeline as a clip, 'solid:#rrggbb' , 'motion:<id>' for a MOTION composition (create_motion_composition) for a generated solid-colour matte, 'adjustment:' for an adjustment clip — no picture of its own, its effects/masks/opacity grade everything below it — or 'text:' for a title, restyled via set_clip_properties' `text`), atFrame (timeline position), sourceInSec/sourceOutSec (the segment of the source to use — align these with transcript words or sceneSamples), and optional trackId (from get_timeline) to target a specific same-kind track — without it clips land on V1/A1 with overwrite semantics. A spare empty track always exists above the highest used one, so V2/A2 is available before anything sits on it. Video sources with audio get a linked audio clip automatically (on the matching A track: V2 ⇒ A2). Returns the created clip ids in order. Frames are in the target timeline's fps.

frapea's BUILT-IN sound-effects pack — synthesized, licence-free, no API key, no download wait. Reach for this FIRST; use search_sounds only for something the pack does not cover (ambience, foley, a specific real-world object). Call with no `ids` to get the catalogue: each sound carries what it sounds like AND when to use it. Call with `ids` to copy them into the project's media pool as `sfx-<id>.wav` at the Library root; each result gives a mediaRef ready for add_clips on an audio track. Adding a sound already in the project is a no-op that returns its existing mediaRef.

Arranges clips into a screen layout by writing ordinary transform + crop values (first clipId → first cell). layout: full | side-by-side | top-bottom | pip-top-left | pip-top-right | pip-bottom-left | pip-bottom-right | grid-2x2 | main-sidebar | three-up. mode 'fit' letterboxes each source in its cell, 'cover' fills the cell and crops the overflow (bias 0–1 picks the surviving part; 0.5/0.5 = centre). Replaces the clips' geometry animation. The clips should overlap in time on different tracks — a layout of clips that never share a frame arranges nothing visible.

Waits (up to ~45 s per call) for the user to resolve a pending Frapea prompt — e.g. the folder-permission dialog another tool just opened. The user cannot see tool calls: BEFORE waiting, always say in your visible reply what to click in the Frapea tab, or they will never know they must act. Returns {status: 'granted' | 'denied' | 'pending'}. On 'pending', repeat the instruction and call again; on 'denied', stop that branch and explain the consequence.

Makes a clip FOLLOW a tracked face — the way to stick a mask, label or blur onto somebody's head. Movement is relative to where the face was first seen, so the clip keeps the framing it already had and only inherits the motion. Pick what to copy with channels, and what happens while the face is missing with gap. Pass trackId: null to unbind.

Navigates the user's Frapea tab back to the project manager and releases this session's headless runtime for the project.

Pairs you with the user's Frapea browser tab. TWO WAYS ROUND. (1) If they already have Frapea open: ask them for the pairing code (the Connect MCP button) and call this with it. (2) If they do not, or you are not sure: call this with NO arguments and you get back a link — send them the link, then call this again with the code it contains every few seconds until it returns a token (they have to accept it in their browser first). Either way you end up with a session token to pass on every later call. Codes are single-use and expire in ten minutes; a token lasts until the user disconnects.

Asks the user to attach a folder of THEIR OWN source footage to the project — opens a dialog in Frapea with a 'Connect folder' button. Only a human can do this (browser security gesture): tell the user IN YOUR REPLY to click it, then call await_user_action with the returned actionId; on 'granted', get_media will list the folder's files. Use when get_media comes back empty and the user has footage somewhere on their disk. NEVER call it for files YOU made or fetched. The project folder's `media/` folder is scanned: any media file placed there joins the pool by itself (within a second, or when the user returns to the tab), with no dialog. `reason` is shown in the dialog word for word — say which footage you need and what you will do with it, so the user knows what to pick.

Creates a MOTION COMPOSITION — an animated graphic (infographic, title, promo end-card, chart…) written as a React TSX component. Read the 'motion' skill (read_skill {topic:'motion'}) for the exact API surface, available fonts/icons and golden examples before writing one. files maps file names to TSX source; the entry must `export default` the component. schema declares the editable props (type string|number|color|boolean|select|mediaRef + default) — they become Inspector controls the user can tweak, so parameterize colors, copy and media slots. The composition is compiled and test-rendered BEFORE saving; on failure the error returns verbatim — fix the code and retry. Returns the mediaRef to place with add_clips. Unless asked for plain, aim well beyond a PowerPoint of titles and bullets — the story, its moments and the look are yours to decide. With UI-like parts (buttons, pills, progress bars, cards), write them in their own file with a partsBoard prop (default false) and look at the board first: preview:true, previewProps:{partsBoard:true} returns one large frame with the reply.

Creates a new timeline in the project and returns its id. Set fps, width and height for the target format (e.g. 1080×1920 @30 for a vertical short). The new timeline starts with one video and one audio track; spare tracks appear automatically as clips land.

MULTICAM angle switch. The angles are ordinary video tracks stacked over each other (sync them beforehand — clips of the same take on V1..Vn); this razors the whole angle stack at `atFrame` and keeps only `chosenTrackId` live from there to the next cut — the other angles' clips are DISABLED, not deleted, so every switch stays re-decidable (cut again with a different angle to change your mind). Linked audio follows the video by default (audioFollowsVideo: false keeps audio untouched — the usual choice when one good mic carries the sound). angleTrackIds defaults to every video track with a clip under the frame. Refused when the chosen track has no clip at the frame.

Downloads one Freesound sound (its HQ MP3 preview, ~128 kbps — plenty for SFX/ambience) into the project's media pool as a virtual file at the Library root (name `freesound-<id>.mp3`). Asynchronous: the first call accepts the download and returns its state; CALL AGAIN with the same arguments to poll — state becomes 'done' with the asset's mediaRef, ready for add_clips onto an audio track. Three files stream at a time and the rest wait their turn, shared with stock video, so state 'queued' is normal and needs no action but polling.

Downloads one Pexels asset into the project's media pool as a virtual file at the Library root (name `pexels-<id>.jpg|.mp4`). Asynchronous: the first call accepts the download and returns its state; CALL AGAIN with the same arguments to poll — state becomes 'done' with the asset's mediaRef, ready for add_clips. maxHeight picks the best video rendition not exceeding it (default: the best available). ASK FOR THE WHOLE SHOT LIST AT ONCE — three files stream at a time and the rest wait their turn, so a batch is queued for you, in the order you asked. State 'queued' is normal and needs no action but polling. ONCE DOWNLOADED, CHECK THE MOMENT YOU MEAN TO USE — not just the first frame. A clip that opens on your subject can cut away, change subject or hold a logo exactly where you planned to trim. Read the scene samples from get_media, or render frames with inspect_media, across the range you intend to cut. See read_skill {topic:'verify'}.

Copy a timeline under a new name — a backup of the cut before a review round changes it, or a variant to try something on. A DEEP copy: new ids for the timeline, its tracks, clips and transitions; links between clips (A/V pairs, caption groups) are kept, inside the copy. Nested timelines are copied too by default (includeNested), so editing a nest in the original does not change the backup; motion compositions and media are referenced, not copied. The original is not touched and stays the active timeline; one undo removes the copy. Returns {timelineId, name, clips, nestedCopies?}.

Generates animated captions from a spoken clip's transcript (the same generator as Frapea's UI): transcript words group into caption text clips over the clip's span, word-timed so the animation follows speech. One undo, and they share a captionGroupId — restyle them all via set_clip_properties with text.wholeCaptionGroup. Needs a transcript (get_transcript tells you). Returns the created clip ids. PLACEMENT: they JOIN the existing caption track whenever they fit in its gaps, and only open a new one when they would overlap (a second language, a second layer). So captioning four narration segments in four calls leaves ONE caption track, not four — no tidying needed afterwards for this. Call manage_tracks {action:'tidy'} at the end of the build for the empty tracks left by everything else. They join the LOWEST caption track, so a picture layer you add ABOVE it after an earlier pass will cover them — captions render behind b-roll rather than looking wrong, which is the hardest kind of mistake to notice. Move the whole lane with manage_tracks {action:'merge'} if you want the captions on top.

Generates an AI voiceover from a script, entirely on-device. The WAV lands in the project's media library as a normal audio asset (with waveform and analysis), the script is stored as asset metadata, and a script-corrected word-level transcript is written so captions work immediately. CALL THIS BEFORE YOU CUT THE PICTURE. A script does not tell you how long it takes to SAY, so a narrated edit assembled first is built on guessed cut points that the real audio then breaks — every section, graphic and transition has to be retimed. Generate the narration, read its word timings with get_transcript, then lay picture against those frames. GENERATE IN SEGMENTS, NOT ONE BLOCK — call this 2-4 times (hook / body / close, or one per section) and lay the parts on the timeline. A single long generation comes out FLAT and evenly paced, everything pressed together at one energy; segmenting gives each part its own intent, puts natural air at the joins, and lets you redo one section instead of all of it. Write for the ear: short sentences, one idea each, LINE BREAKS between them — the strongest signal the model has to actually stop. See read_skill {topic:'voiceover'}. Async: returns {jobId, fileName} — fileName is the audio file's final name in the project's media/ folder, decided up front (WAV for local providers, MP3 for ElevenLabs). The job runs on the analysis MASTER tab and its request is persisted on disk: it survives tab closes and browser restarts, and get_job_status answers from disk even in a fresh session. Poll until done; the finished job also carries the asset's mediaRef. Voices/languages come from list_voiceover_models; a local model must be downloaded first (phase 'ready'), otherwise this returns PERMISSION_REQUIRED. PREMIUM (provider 'elevenlabs' — SPENDS THE USER'S CREDITS): check list_voiceover_models' `elevenlabs` status object first. Structured refusals you must handle: NO_API_KEY (no key stored — ask the USER to add their key in the Voiceover generator's ElevenLabs onboarding; never ask for, or pass, the key yourself), INSUFFICIENT_CREDITS (carries needed vs remaining — shorten the script or ask the user), CONSENT_REQUIRED (paid features are set to 'ask': a consent dialog with the cost is now open in Frapea and the error carries an actionId — TELL THE USER in your visible reply to confirm it, call await_user_action with that actionId, and on 'granted' retry this call with consentActionId set to it; 'denied' means drop it). AUDIO TAGS (ElevenLabs): inline bracket tags like [calm], [excited], [whispers] direct the delivery ONLY on models whose list_voiceover_models entry says supportsAudioTags: true (the v3 family). Every other model reads them ALOUD — so only write tags when the chosen model supports them. As a safety net, tags sent to a non-supporting model are stripped before synthesis and the success message says so.

How far the project's media analysis has got. Analysis runs BY ITSELF whenever media is connected — never start it. Returns percent (0-100) — which reaching 100 only means nothing is OUTSTANDING, so always read `failed` beside it and call retry_analysis on anything that failed — done/total/pending item counts, runningItem {mediaRef, kind, progress}, etaSeconds (null until something has finished), blockedOnModels, and models {visual, speech}. Poll this while you wait; when blockedOnModels is above zero, call request_ai_models.

Prepare a BUG REPORT for the user: it opens in Frapea (Problems → Create bug report) for them to read and save. The report itself is NOT returned to you — it would leave the user's machine before they have seen it. You get a summary: open problems, media counts, and how many of the user's own words remain in composition code. It contains versions, problems with engine details, the project's structure with names and texts replaced, provider files by public id, the user's own files only as encoding fingerprints, and motion code. timelineIds limits it to those timelines and what they nest.

Status of a background job started by a job-returning tool (generate_voiceover, verify_timeline, measure_loudness): {state: queued|running|done|failed|cancelled, progress: 0-1, detail, ...result fields such as mediaRef when done; verify_timeline, measure_loudness puts its report in `result`}. Poll every few seconds while state is 'queued' or 'running' (queued = waiting for the analysis master tab, which loads the models once for the whole browser). For analysis started by start_transcription or start_visual_indexing, poll get_analysis_status instead.

Lists the project's media library: mediaRef (the id clips use), name, kind (video/audio/image), durationSeconds, resolution, and sceneSamples — timestamps (seconds) where the visual indexer detected distinct moments. sceneSamples is a ready-made shot map: inspect or cut at those times first. Each asset also carries visualIndex and transcript objects: {phase: 'ready'|'pending'|'queued'|'running'|'failed'|'no-audio'|'disabled', error, attempts, sealed}. Analysis starts by itself and retries a failure a few times before sealing it. If a phase is 'failed', read `error` and decide: call retry_analysis to run just that one analysis again, or leave it. If 'disabled', call request_ai_models. Video and audio assets also carry beats: {phase: 'ready'|'pending'|'failed'|'none', bpm, confidence}. These come from an ONSET detector (spectral flux): every energy rise — drum hits, plucks, syllables — not bars or downbeats. bpm is the tempo of those onsets, or null when there is no steady pulse (confidence 0–1 below the detector's 0.3, or too few onsets) — speech has onsets and no bpm. 'none' = no audio to analyze. Pass mediaRef to get ONE asset with beats.onsetsSec (source seconds, the first 2000; truncated + onsetCount say when there are more) — map them through the clip's trim and speed yourself.

Returns a motion composition's manifest and full TSX sources.

What has gone WRONG in the user's editor — renders that failed or stopped, footage that would not decode, chunks that could not be prepared for playback, exports that failed — each with WHERE (the timeline and frames, the source and its own frames, the footage and frame the engine could not decode) and the reason in the engine's own words. You also hear about new problems for free: every other tool result carries a `problems` list of what changed since your last call. Call this for the whole history, or after a `moreProblems` count. A problem that is open on a timeline means that timeline does NOT play or export correctly there, whatever inspect_timeline showed: fix the named layer (e.g. update_motion_composition) before calling the work done. `state` is open | fixed | superseded (the thing it was about was edited and not yet re-rendered — NOT proof it works) | dismissed.

Lists the projects Frapea knows on this machine: id, name, lastOpenedAt, and for a project stored in a folder on disk its `folder` (the directory's name — files you write into its `media/` subfolder join the pool by themselves). Call first to find the projectId every other tool needs. A project the browser has no permission for will surface PERMISSION_REQUIRED when used — relay the returned instruction to the user, then retry. ALSO RETURNS `providers` — Pexels (stock video/photos), Freesound (music, ambience, SFX) and ElevenLabs (voiceover), each 'connected' | 'checking' | 'missing' | 'invalid'. These are the user's OWN API keys and they cannot be set from here. When any is missing, `providersAdvice` is present and carries the exact thing to say. SAY IT ONCE, in your first visible reply of the session, and then get on with the work you can do — the user cannot see tool calls, so silence here reads as the editor simply not being able to fetch b-roll or narrate a script. When everything is connected there is no `providersAdvice` and you say NOTHING about it: telling someone to connect what they already connected is nagging. SAY IT AGAIN ONLY WHEN IT IS SOMETHING NEW. The rule is one general nudge at the start, and after that only a CONCRETE consequence for the piece in hand — "I can cut what you shot, but with Pexels connected I'd cover the gaps at 0:14 and 0:37 instead of holding on a static frame" is new information and worth saying once before you build; "remember you have no Pexels key" is nagging. Never as a reminder, never twice for the same reason. SEPARATELY: when a CALL actually fails with PERMISSION_REQUIRED, always relay that error — it is a thing that just went wrong, not a repeat of the nudge, and swallowing it makes a failed fetch look like a choice you made.

Reads a project's timeline structure: timelines (id, name, fps, resolution) with tracks and clips. THE USER CAN COPY A CLIP'S ID from the editor's right-click menu, so a bare UUID in their message names exactly one clip — call this with NO timelineId to search every timeline and match `clip.id`, rather than guessing from a name. Clips carry id, name, mediaRef, kind, startFrame, endFrame (exclusive), trimStartFrame and linkGroupId (linked A/V pairs share it), plus `look` (opacity, blend, transform, crop, keyframes) whenever any of it differs from the defaults — an untouched clip has no `look` key. Omit timelineId to get every timeline; frames are in that timeline's fps.

Returns the word-level speech transcript of one media asset: language and words [{text, startSec, endSec}]. Only assets the user has transcribed have one — NOT_FOUND means ask the user to run Transcribe on the clip in Frapea. Use the timestamps to align cuts with what is being said.

Ripple-inserts media segments: everything at/after atFrame shifts right to make room (clips under the cut split). Same clip shape as add_clips. Use to splice a shot INTO an existing sequence without overwriting.

Looks INSIDE a source asset. Returns ONE contact sheet — every requested moment as a labelled cell in a single grid image — plus a per-cell fingerprint. Pass atSec (max 12 timestamps, seconds); combine with get_media's sceneSamples for an instant storyboard of the whole file. Shows the RAW source; use inspect_timeline for the composed cut. START WITH pixels:false WHEN YOU ARE ONLY CHECKING FOR EMPTINESS. The fingerprint answers 'is this black / flat / did it decode' for no image cost — nonBlankRatio near 0 is an empty or black frame, colorRange near 0 is a featureless one. Ask for pixels when you must judge WHAT is in the frame: subject, framing, legibility.

WATCHES the cut. Composites every video layer at each requested frame and returns ONE contact sheet — the frames as labelled cells of a single grid image — plus, per cell, the visibleClipIds that contributed pixels and a fingerprint. Pass frames (timeline frame numbers, max 12). Use after every edit: this is the only tool that shows what actually PLAYS, as opposed to what get_timeline says is arranged. SAMPLE THE CUT POINTS AND THE MIDDLE OF EACH SECTION, not only frame 0. An entrance animation that never resolves, a graphic clipped by its own box, an offline clip rendering black — all look perfect in get_timeline and are obvious here. USE pixels:false FOR A CHEAP SWEEP. visibleClipIds plus nonBlankRatio confirms 'the right clip is in every slot and none of them is black' across a dozen positions with no image cost; spend pixels on the few frames where you must judge typography, framing or timing.

Lists face-tracking runs and the people each one found: faceId, name, frame range, how many frames it was actually visible for, and how many gaps it has. Sample data is deliberately not returned. Use this to poll a queued run and to pick a faceId for bind_face.

Lists the project's motion compositions: id, mediaRef (for add_clips), name, size, fps, duration and the editable props schema.

The text-to-speech providers Frapea can generate voiceovers with. Returns providers[]: the FREE in-browser models straight from the TTS registry ({id, label, languages, voices: [{id, label, language, gender, grade}], modelAssets, state: {phase: needs-download|downloading|ready|error}}), plus the PREMIUM cloud provider flagged premium: true ({id: 'elevenlabs', label, premium, voices: [{id, label}], models: [{id, label, creditsPerCharacter, supportsAudioTags, note}]}, voices/models empty until a key is stored). supportsAudioTags tells you whether that model understands inline audio tags like [calm], [laughs], [whispers] as delivery directions (only the v3 family does); on every other model such tags are READ ALOUD, so send those models plain text — each model's `note` restates its rule. Also returns a top-level status object `elevenlabs`: {provider: 'elevenlabs', hasApiKey, credits: {used, limit, remaining} | null, consent: 'ask' | 'always'} — READ IT BEFORE choosing the premium provider, so you know whether a key exists, what the account can afford and whether a consent dialog will be involved. A free provider whose phase is 'needs-download' cannot generate until the user enables it in the Generators panel (Voiceover) — there is no tool to force the download; ask the user.

Edits a clip's effect stack (ordered, per-effect enable). action: 'add' (type; returns the new effect's id) | 'remove' (effectId) | 'move' (effectId + toIndex — the order IS the render order) | 'toggle' (effectId + enabled) | 'set' (effectId + params patch, clamped to the effect's ranges) | 'curve' (effectId + curve + points). Types: 'colorBasic' (Lumetri Basic Correction-compatible: temperature/tint -100..100, exposure -5..5 stops, contrast/highlights/shadows/whites/blacks -100..100, saturation 0..300, vibrance -100..100), 'colorWheels' (Resolve lift/gamma/gain/offset — {group}Master/{group}R/G/B; lift/gamma 0 neutral, gain 1, offset printer points 25 neutral), 'lut' (.cube from the project library — set `resource`), 'curves' (tone curves), 'hueCurves' (hue-vs-hue/sat/lum). Also 'motionBlur' — NOT the shutter blur: a DIRECTIONAL blur filter (angle 0–360, distance 0–100) that smears every frame the same way whether the clip moves or not (a streak look, a speed effect on footage). For blur that follows the clip's own keyframed movement, use set_clip_properties motionBlur instead.

Full control of the background work queue. action 'list' returns every queued/running item across all projects (analyses, proxies, beats) plus the visible task rows (downloads, voiceover jobs, failures) with a cancellable flag. action 'cancel' stops ONE kind of work for ONE asset — queued or already running (a running visual index stops at the next sample and keeps its checkpoint). The asset then shows an incomplete-analysis badge; retry_analysis or the tile's retry button runs it again. kind: index | transcribe | stabilize | denoise | proxy | beats.

Track operations on a timeline. action: 'mute' | 'unmute' | 'solo' | 'unsolo' (audio only; any solo silences non-solo tracks) | 'hide' | 'show' | 'rename' (needs name) | 'duck' | 'unduck' (audio only: dip this track under speech detected on the other audio tracks — transcript word timings when available, energy VAD otherwise) | 'merge' (needs intoTrackId: moves every clip of trackId onto another track of the same kind; refuses the WHOLE move if any clip would overlap, or if the source carries transitions) | 'delete' (an EMPTY track; refuses one that holds clips — deleting those is remove_clips, by name) | 'tidy' (NO trackId: drops the empty leftover tracks and renumbers). Track ids come from get_timeline. TIDYING UP, AND WHY ORDER MATTERS. Track counts only ratchet up while you build: a spare appears above whatever you fill, and the video and audio counts are kept EQUAL. That last rule is the one that surprises people — five video tracks force five audio tracks, however empty. So: 1. 'merge' the tracks that hold clips but need not be separate (captions from several passes, b-roll that never overlaps). This is what actually lowers the count. 2. 'tidy' once at the end. It settles both sides on one spare above the taller one and reports removedTracks. 'delete' on a single empty track usually reports removedTracks: 0 — the invariant puts it straight back to keep the counts equal. That is not a failure, it is why 'tidy' exists. Neither ever touches a clip, and both keep an empty track you renamed, muted, hid, soloed, unlocked or set to duck.

Transitions on a track's cuts OR lone clip edges (a free head/tail takes a single-sided fade from/to nothing; its alignment is forced). action: 'add' (trackId + atFrame = the cut where one clip ends and the next starts, + type; optional durationFrames (default 30, clamped to the clips' source handles), alignment centered|start|end, params) | 'remove' (transitionId) | 'update' (transitionId + any of type/durationFrames/alignment/params). Types: crossDissolve, dipToBlack, dipToWhite, wipe (params.angle deg, params.softness 0-100), slide (params.angle), push (params.angle). Both clips need source material past the cut — a transition that cannot fit is refused; edits that break the cut remove it. Returns {ok, transitionId?}.

MEASURE THE MIX — you cannot hear it; this is how you check it. Mixes the timeline's audio offline exactly as an export does (every audio track, nested timelines, fades, volume keyframes, clip audio effects, ducking, mute/solo) and measures it with EBU R128 / ITU-R BS.1770 — the same meter export loudness normalization uses. The mix is measured AS MIXED: an export's loudnessTargetLufs would then shift all of it by one gain. Returns {jobId}; poll get_job_status every few seconds. When done, `result` is {durationSec, integratedLufs, truePeakDbtp (4× oversampled, BS.1770-4), samplePeakDbfs, loudnessRangeLu, maxShortTermLufs, windows: [{fromSec, toSec, integratedLufs, maxShortTermLufs}] one per windowSec, sections?: [...same per requested section], speech: {source: transcript|none, speechSec, gapSec, mixDuringSpeechLufs, mixInGapsLufs, transcribedClips, untranscribedClips}, tracks: [{trackId, name, carriesSpeech, integratedLufs, samplePeakDbfs, duringSpeechLufs, inGapsLufs}]} (true peak is measured on the whole mix). SPEECH comes from transcript word timings mapped onto the timeline (trims, nests, speed): a GAP is a stretch of at least 0.6 s with no word within 0.25 s of it. `inGapsLufs` of a music/ambience track is the bed's level where nobody talks; compare it with the speech track's `duringSpeechLufs`. Without a transcript (speech.source 'none') the speech/gap fields are null — run start_transcription on the dialogue first. Integrated values are gated (−70 LUFS absolute, −10 LU relative), so silence does not drag them down; null means too short (under 0.4 s) or silent. Needs the project open in the editor tab you are paired with (PROJECT_NOT_OPEN otherwise).

Joins two or more faces from one run into a single person — for somebody who walked out of the shot and came back. REFUSED when any two of them appear on the same frame, since they cannot then be one person. The earliest face's id and name survive, so existing bindings keep working, and the time between sightings stays a gap for the clip's gap policy to handle.

Moves clips to new positions (linked partners follow automatically). Each move: clipId, toStartFrame, optional toTrackId (same-kind track). Overwrite semantics — landing on another clip trims it like a drag-drop.

Creates a project. By default it is created in the browser's own storage, which needs no permission and no click: the project id comes back immediately and you can start working. The user can turn it into a real folder on their disk whenever they want (Cmd/Ctrl+S, or File → Save to a Folder) without losing anything. Pass storage:'folder' only when the user asked for a folder up front — that needs a browser security gesture, so it opens a dialog in Frapea with a 'Choose folder' button, returns an actionId, and you must tell the user IN YOUR REPLY to click it, then call await_user_action; on 'granted' the project appears in get_projects.

Navigates the user's Frapea tab to a project so they can WATCH the edit live. Not required for editing — every tool works headlessly; use this for show-your-work moments.

Pool file operations. action: 'rename' renames the real file on disk (needs newName; clips referencing the old name go offline). 'delete' UNLINKS the file from the media pool — the real file on disk is NOT touched; clips using it go offline until it is added again.

THE MEDIA POOL'S SHAPE — bins (the pool's folders) and what sits in them. Use it to tidy a project: group the footage, the voice-over, the music and the graphics instead of leaving eighty files at the root. Bins are VIRTUAL and live in the project, not on disk. Moving an item between bins NEVER touches the file or its mediaRef, so it cannot break a clip's link or make the app ask the user to relocate anything — organise freely. actions: 'list' returns the bin tree and every item with the bin it shows in (call it first — you need the ids). 'create_bin' {name, parentBinId?}. 'rename_bin' {binId, name}. 'move_bin' {binId, toBinId} nests one bin inside another. 'delete_bin' {binId} — its children and items reparent, nothing is lost, no file is deleted. 'move_items' {itemKeys[], toBinId} moves files, timelines and motion compositions alike. Item keys: a media file's mediaRef, 'timeline:<id>', 'motion:<id>'. The Library root is the empty-string bin id. Bins that MIRROR a real folder (mirrorsFolder: true) can be RENAMED — a mirror is matched by the folder it tracks, not by its name — but they cannot be moved or deleted, because their place in the tree follows the folder. Move the items out instead, or make your own bins alongside.

Returns Frapea's agent guide: editing conventions, tool workflow tips, the current feature surface, and the CREATIVE MANDATE — how far to run with a terse brief (source stock footage, sounds, voiceover, motion graphics and deliver a finished video vs. when to hand control back to the user). Read once at the start of an editing session. MOTION topics: 'motion' is the API surface (imports, fonts, props schema, the parts-board-first golden example) and routes to the rest — read before create_motion_composition; 'motion-examples' (two finished pieces that differ in look); 'motion-craft' (anticipation, overshoot, stagger, the tells of machine-made motion — read with 'motion', always); 'motion-text' (size floors, fitting copy that cannot clip, per-word animation, annotations); 'motion-effects' (grain, duotone, light leaks, blurs, glow); 'motion-transitions' (TransitionSeries, presentations, cut overlays); 'motion-sound' (where sound lives, the built-in SFX pack, placement and levels); 'motion-review' (the same loop as 'review', from a motion job). The CRAFT guides teach how to actually cut: 'core' (rhythm, b-roll, audio hierarchy, restraint — read first), then the format you are making — 'shorts' | 'longform' | 'podcast' | 'tutorial' | 'montage'. Also 'voiceover' (read BEFORE generate_voiceover: narration is generated in segments, not one block) and 'verify' (how to prove the cut is right — read before every export), and 'review' (for any piece with an audience: an opponent argues with your rendered frames in rounds until it is finished — read before you start).

Deletes clips (lift — leaves a gap). Linked partners are NOT removed automatically; pass both ids of a pair to delete it whole. To close the gap too, use ripple_delete_ranges instead.

Cuts silent gaps inside a clip: any pause between transcribed words longer than minGapSec becomes a ripple delete (a small margin is kept so speech never clips). Typical minGapSec: 0.6 tight, 1.0 balanced, 1.5 loose.

Text-based editing: cuts the given transcript words OUT of a clip as ripple deletes (everything after each cut shifts left). wordIndices are positions in get_transcript's words array for the clip's media. Contiguous indices merge into single cuts. The go-to tool for removing filler words or tightening an interview.

Asks the user to enable Frapea's on-device AI analyzers (picture search, speech). Opens a dialog in every Frapea tab and returns an actionId — tell the user IN YOUR REPLY to confirm it, then call await_user_action and poll get_analysis_status until models are enabled. Call this when get_media reports 'needs-models'.

Runs ONE analysis of ONE asset again — picture indexing or speech transcription — clearing whatever failure was recorded for it. Use when get_media shows phase 'failed' and the error looks worth another go, or when you want a fresh result for a file that changed meaning. Returns the project's analysis status.

Extracts [fromFrame, toFrame) on one track and shifts everything after it left to close the gap (sync-locked tracks follow). The go-to tool for tightening a cut or removing a bad take.

Semantic search over the project's media. Visual scope matches what is IN the picture (describe a shot: 'aerial city at night'); spoken scope matches transcribed speech (exact words, ranked first, with a snippet). Hits: {mediaRef, timeSec (best moment), fromSec, toSec, score, scope}. Requires the user to have downloaded the picture analyzer — otherwise PERMISSION_REQUIRED tells you what to relay.