add_music
ChatGPTWrite music, beat timing and a ducked audio mix. Requires sections or direction; direction alone requires a stored timeline. Sound effects use the plan's sfx beats. Takes the project lock and returns task_id; get_task reports completion. Repeats start another priced run. Billing: free.
add_sound_effects
ChatGPTGenerate sound effects at the points you choose on the finished cut's timeline (at_ms) and mix them with the voice and any music. Send 1 to 12 hits in one call: each call replaces the project's previous sound-effect track. A hit that starts past the end of the cut is refused, not retried. Takes the project lock and returns task_id; get_task reports completion. Repeats start another priced run. Billing: free.
build_captions
ChatGPTBake the caption layer from the subtitle pattern the caller picks into compositions/narrator_captions.html and transcript-src.json. A host project's result carries host_div for index.html and files ({path, sha256}) to commit, adding those entries when you name expected_files, or they stay staged. position, placement, hide_intervals, caption_windows, emphasis_word_ids, font and accent are optional; one model call fills the pattern and emphasis left out. Billing: paid, priced by source length and lane; get_pricing states the amount.
cancel_task
ChatGPTCancel one active orchestration task, queued or running. Cancellation is graceful rather than immediate: the worker finishes its terminal bookkeeping first, so status reaches cancelled some time after the request is accepted. A task holds its project's lock until it reaches a terminal status, and while it is held every tool that starts a stage or an edit on the project refuses with the active task id; a task nobody cancels holds the lock until it ends on its own. A paid task you cancel keeps what it had already spent, never more than the amount it held, and releases the rest. Billing: free.
commit_workspace
ChatGPTJudge every file staged for one project and accept all or refuse all; only this call moves a staged file into the workspace a render reads. A composition is refused for external scripts, inline handlers, javascript: URLs, network calls in executable script code, or more than 64 KiB (4 MiB for compositions/narrator_captions.html), because it is served into the project preview beside the page. A sub-composition is also refused when it carries no <template> (missing_composition_template), when its window.__timelines[id] = tl assignment sits after </template> (timeline_outside_template), when it registers a key no host in the file carries (timeline_id_mismatch) - only template content is cloned, so that timeline never registers and the export blocks on it - or when a script binds a literal [data-composition-id=id] selector with a class the host does not carry or :not(.visual-host), or sends a :scope > child query to that root, because the export mount strips the inner root's data-* attributes and only the host carries the id, so that root reads null and the script throws before it registers. A composition's media, stylesheet and mount references (img, video, audio, source, link, data-composition-src) are read from the project root (public/..., styles/..., compositions/...), never from compositions/, because the editor mounts the template into index.html; a ../ path there names a file above the root and is refused as missing_reference. Only a composition's script src is read from its own file (../public/vendor/gsap.min.js), as the export's page load reads it, and a url() inside a .css file is read from that stylesheet. Every file a document or a live style names must be in the workspace or in the same commit, or the round is refused as missing_reference instead of exporting without it; a @font-face src is the exception, since the browser falls back when a face is absent. data-start, data-duration and data-media-start are seconds, and a value that is not a finite number in range (start and media offset at or above 0, duration above 0) is refused as invalid_timing rather than failing the export later. A host (data-composition-src) written inside another host is refused as nested_composition_host, since the outer composition fills its host and the inner one never registers; close each host's </div> before the next host opens. An <iframe srcdoc> that embeds a whole HTML document, or any other attribute value, inline <script>, <style> or HTML comment holding the text </body>, in index.html or inside the <template> of a composition this round writes or newly mounts, is refused as embedded_body_close: the export inserts its frame-capture script before the first </body> of the page it builds, so that one captures it and the render fails; mount such a slide as a composition with a data-composition-src host instead of an iframe. A sub-composition <style> the export's CSS parser rejects (a missing semicolon, an unbalanced brace, a stylesheet cut off mid-rule) is refused as composition_css_syntax with the parser's reason, line and column, when this round writes that composition or newly mounts it, since the preview tolerates it but the export compiles every mounted composition stylesheet and fails. An inline <script> of a document this round writes, or a .js file it writes, that the browser cannot parse (a script collapsed onto one line so a // comment swallows the rest, an unbalanced bracket, placeholder text pasted into the code) is refused as composition_script_syntax with the reason, line and column, since a script that does not parse stops every script after it in the export and nothing registers its timeline. An index.html whose root declares a canvas larger than 1080p (data-width/data-height with a long edge past 1920 or a short edge past 1080) is refused as composition_canvas_too_large, since the export is at most 1080p; author at 1080x1920, 1920x1080 or 1080x1080. An index that keeps the canvas the workspace already has is accepted. The caption layer, the composition declaring data-composition-id narrator-captions, is accepted o…
enhance_voice
ChatGPTEnable voice enhancement for an editable project, or disable it; disabling reverts to the original audio. A separate task the user chooses explicitly: creation never runs it. Returns task_id to poll with get_task. Billing: free.
find_broll
ChatGPTFind b-roll images for this project: one image search per query, then a vision pass picks the best candidate for each, and the reply carries task_id to poll with get_task. List a query in cutout_queries when its moment needs the subject cut out on a transparent background; one not also in queries is refused. Billing: free.
finish_project
ChatGPTMark an externally created, settled editable preview finished. Returns task_id, or success without a task if already finished. Refuses a busy or failed project or one without a settled editable preview. Does not assess editing quality, export or add treatments; the project remains editable. page_url opens the project in the product app, with its navigation and without the AIP chat, and is the durable link to give a person in the final reply. agent_page_url opens the editor view for embedded panels; it is not the final delivery link. Billing: free.
frame_speaker
ChatGPTMeasure the speaker per window (cut seconds) and, with matte true, cut them out. The task result gives each window its on-camera share, face and head band, and for a slot the object_position that keeps the head whole and fits; a matted window adds its pop geometry (sink, card edge, headroom) and render-engine/public/speaker_<slug>.webm, staged for your next commit_workspace (media_start 0). Never writes your tree. Refused before compute: no transcript; a matte without a slot, off camera, unmeasured, or across a cut; over 8 windows; a matte over 30 s a window or 60 s a call. Billing: free.
generate_scene_restyles
ChatGPTGenerate a scene restyle cover for each window and write it into the project workspace; a timeline this server built also gets it placed, one you authored does not. Up to 25 windows per call, each start_ms to end_ms spanning 4-10s of output time, non-overlapping; a window out of bounds is refused at no cost. Covers a submitted visual plan reserved are bought after your windows while the hold has room; the rest stay pending. A window already covered is not bought again; an identical window addresses the one it made. Each verdict in get_task.result carries the cover's workspace path, whether the timeline already carries it, how many generations this call was priced for (generations, billed_generated_seconds), and for a refused window its class, the provider's reason, and whether it could ever differ; that window keeps no cover, plays the raw footage, and is terminal: a re-run never re-attempts it. A content filter refusing a recognizable individual refuses every window of that speaker. A window may name a preset from list_motion_assets.scene_restyle; a preset outside this project's catalogue is refused at no cost. While the project's Scene restyle switch is off the call is refused at no cost unless enable_skill_quote carries the user's literal words asking for it, which turns the switch on for this project first. Billing: paid, 6 credits per started second of video generated, each window once and twice when the identity check rejects its first try, whether or not it ends with a cover, and never for a generation returned in the wrong frame shape; the hold is twice every window sent, and a window already covered, or refused before generating, costs nothing; get_pricing states the amount.
get_chat_history
ChatGPTRead one stable page of a project's canonical chat transcript. The cursor is exclusive. Forward order starts at the oldest message and can enumerate the complete history; backward order starts at the newest. next_cursor is null on the last page. Billing: free.
get_editing_preset
ChatGPTThe editing preset saved on the calling account: the answers it keeps as defaults for a new project. Returns editing_preset, null when nothing is saved. The preset is a record of what the user chose; it admits no stage, changes no price, and no other tool on this server reads it. caption_style=no-caption means skip caption generation and leave the caption layer unmounted, not a pattern. Billing: free.
get_export_url
ChatGPTRead the download link for the project's latest delivered export. state ready carries download_url and download_expires_at; deferred means a live turn still owns the project; not_done means an export has not delivered; no_export means none was requested. Render progress is get_task's, never here. download_url is a one-use credential minted for the CALLING account (not the owner, which differs on a staff read) and signs whoever opens it into that account. A ready reply lists dropped_visuals, moments the render shipped without because their setup script threw, and what the workspace looks like NOW: current_workspace_digest, staged_files_present and warnings. Those describe the workspace, not this file, which may have rendered an earlier one; current_workspace_digest is null when it could not be read. page_url opens the project in the product app, with its navigation and without the AIP chat, and is the durable link to give a person in the final reply. agent_page_url opens the editor view for embedded panels; it is not the final delivery link. Billing: free.
get_motion_asset
ChatGPTRead one package by motion_asset_id and version from the registry: the pinned manifest, the entry source per supported orientation, and its preview. Reads no project. A version the registry does not hold is refused. Billing: free.
get_pricing
ChatGPTRead what every capability costs, in credits. Each row carries band (free / negligible / paid / expensive), credits per run, and available; a per_unit row also charges credits_per_unit per unit consumed, up to max_units. On a row without per_unit, credits is the whole price per run; read per_unit before quoting a row that has one, since such rows carry credits 0 and quoting that 0 calls a paid step free. billing.metered says whether credits move here; free_usage says how much free usage is left; project_id adds that project's spend. available=false rows are not callable. Billing: free.
get_project
ChatGPTRead project phase, status, active task, optional preferences, and editable readiness. include_spend adds ledger spend; null means unavailable. page_url opens the project in the product app, with its navigation and without the AIP chat, and is the durable link to give a person in the final reply. agent_page_url opens the editor view for embedded panels; it is not the final delivery link. Billing: free.
get_project_asset
ChatGPTRead one project asset: a registered upload by asset_id, handle or @asset/handle, else a project Asset Library entry, including generated and Motion library copies, by asset_id or handle. mode inspect (default) returns metadata and native image content for PNG, JPEG or WebP up to 3 MiB. mode download returns metadata and a download_url valid for 15 minutes, including larger or other media. A registered upload path is not a render URL; an Asset Library entry carries its origin, and its path is the workspace file a composition loads. Billing: free.
get_task
ChatGPTRead the live state of the exact task a mutation on this surface returned. Terminal exactly when status is one of done | failed | cancelled; settled only then. succeeded is the outcome to branch on and done is only the lifecycle: a commit_workspace round that refused every file is done with succeeded false, and answers that round in result (outcome, accepted, issues, next_action). failed carries error, and a failed export's result names what to repair: result.issues (composition_id, path, code, cause) or result.planner (reason, duration_s, max_duration_s, plus file and check when it could not read the tree). Beside it, failure carries a stable code and a repair hint where one is known. Show failure.message to the user; error is bounded technical diagnostic text for repair, not user-facing copy. retryable: false means this edit can only fail the same way, null means retryability is unknown. stage, heartbeat_at, attempt and typical_duration_seconds describe progress. deferred is an export parked behind a live turn: it starts when that turn ends. Every reply carries page_url, the durable project page: the link a person can be given. wait_seconds holds the call open until the task settles or its status changes; next_data.poll_args is that held call again: one call per wait, no sleep. A terminal task that took a paid hold also carries billing: credits_held is the ceiling it was quoted, and once settled is true, credits_charged is what it actually took and billable_units what that paid for where the price is per unit. A task turns terminal before its charge settles, so settled false means the amount is not known yet and the next poll carries it; billing absent means no user-paid hold. A project's run_status is not this task's outcome. Billing: free.
get_transcript
ChatGPTRead one project's transcript. detail segments is one line per segment with its text and bounds, a tenth of the bytes: read it first to plan a cut. detail words (default) is every word with its times, what captions are built from. src__ms is a position in the SOURCE media (null when none); timeline__ms is on the CUT timeline. state ready is the whole document; invalid failed its checks (captions.issues says how); pending has no words yet. state deleted was cut; hidden is not drawn. One call is one page: pass page.next_cursor back as cursor, stitch per page.merge_rule. Billing: free.
get_view_url
ChatGPTMint a single-use browser URL for one readable project. url works exactly once: it starts an agent session and redirects to the project page in agent mode, and url_expires_at is when an unused link dies. A session opened under an OAuth authorization lives as long as that authorization (up to about 30 days) and ends the moment it is revoked; any other caller gets about 8 hours. agent_page_url is the durable panel link and reuses the same browser cookie, so it needs no second ticket. Billing: free.
get_workspace_file
ChatGPTRead one file out of a project's render-engine workspace: what is deployed, not what you believe you uploaded. A text source comes back inline in content as utf-8, whole, or the window offset and max_chars name (the reply carries chars and next_offset), so a large file is sampled, not swallowed. Anything else, and any file over 524288 bytes, comes back as download_url with download_expires_at and content null. Readable paths: render-engine/index.html, render-engine/compositions/, render-engine/public/, render-engine/styles/, render-engine/fonts/. Missing answers 404. Read-only. Billing: free.
list_motion_assets
ChatGPTList the Motion library packages with this project's placement state for each, read at project_revision. An entry carries its inputs (a video input that needs a speaker cutout states its slot) and an action: add, chooseInput, chat or unavailable, with the reason and the one call that acts. With at_ms (and the duration_ms you will place with), a cutout package's next carries the complete frame_speaker calls for that effect window, one matte window per kept span. scene_restyle lists the presets this project is served and whether its restyle switch is on. Billing: free.
list_project_assets
ChatGPTList registered project uploads available to composer @asset/handle mentions. Returns asset_id, handle, kind, path and media metadata, without media bytes or download URLs. The source video and other Asset Library entries, such as generated or Motion library media, are not listed; get_project_asset still reads such an entry by asset_id or handle. Billing: free.
list_projects
ChatGPTList your own projects, newest first, one page per call. total counts every project, and next_cursor is null on the last page. An item's progress is not authoritative: get_project is. page_url opens the project in the product app, with its navigation and without the AIP chat, and is the durable link to give a person in the final reply. agent_page_url opens the editor view for embedded panels; it is not the final delivery link. Billing: free.
list_sound_effects
ChatGPTList the sound effect library with each sound's measurements: duration_ms, onset_ms, peak_ms (its loudest moment), hits and hits_ms (where each hit lands, ms from the sound's start), phone (audible on a phone speaker), tone_gate, use_when and suggested_gain_db (the level that sits it under speech), plus this project's project_revision and cut_end_ms. family narrows to one family; query matches every term against names and descriptions. Pages by limit and cursor. Billing: free.
list_versions
ChatGPTList one project's saved editable versions and which of them is live, one page per call in stored order; pass next_cursor back as cursor for the rest. Only a version_id returned here names a version: restore_version answers 404 for any other value. Read-only. Billing: free.
list_workspace
ChatGPTList what one project's render-engine workspace holds, one page per call; pass next_cursor back as cursor. Each file carries path, size in bytes, and md5, the base64 MD5 the store holds (null means unknown). The listing is recursive under path_prefix and covers only render-engine/index.html, render-engine/compositions/, render-engine/public/, render-engine/styles/, render-engine/fonts/. digest is one token over the whole face on every page, whatever path_prefix or cursor; it changes on any write under the face, the editor's included, and is what base_digest is compared against. A cursor is bound to the digest it was read under: a continuation after the face changed is refused with 409 workspace_changed and paging restarts from the first page. staged (what you uploaded that no commit has accepted, by destination; in no render, visible to nobody) and last_promote (the latest commit_workspace round, which the next commit replaces; a commit's own verdict is get_task's result, null before any round) come on the first page only. Billing: free.
materialize_motion_asset
ChatGPTPlace one certified package on this project's timeline at timeline_in_ms, the placement the browser's Add runs. expected_project_revision must equal the revision list_motion_assets returned; another is refused with the current one. A media input is a registered asset id or, for a video input, ordered staged segments {path, sha256, start_ms, end_ms, geometry} from frame_speaker, verified before they bind. The same idempotency_key replays the first receipt. Refused: an unknown or retired package, a required input without a value, a project under a task. Billing: free.
open_aip_app
ChatGPTOpens the AI Producer sidebar app with published video examples. Returns title, message and next_data. Takes no arguments; opening it creates nothing. When the user then attaches a video and asks for an edit, follow the AI Producer plugin's aip skill: it starts the AI Producer project with these tools. Billing: free.
place_sound_effects
ChatGPTPlace library sounds on this project's cut in one undoable edit. Each hit names a sound_id, the at_ms its anchor (first_hit, peak or onset) lands on, and an optional gain_db; a sound is trimmed to the cut's edges, and a hit past the end is refused in refused. replace removes the effects this tool or the generated stage placed before, keeping a person's own. expected_project_revision and expected_library_version must equal list_sound_effects'. The same idempotency_key replays. Billing: free.
present_choices
ChatGPTDisplay a question as an interactive card, with a text fallback for hosts without card support. Returns the question and options, not a user answer; no project setting changes. Component ids: visual_style, caption_style, rough_cut_boldness, editing_preference_example, editing_preference, bgm_enabled, sfx_enabled. project_id associates the card with an accessible project. agent_components supplies custom questions. No task or project lock. Billing: free.
record_choices
ChatGPTStore one decision card's selections. The latest answer replaces that card's previous answer; [] records an empty selection. Returns choices and pending_choices. The archive does not affect execution, pricing or stage admission. Billing: free.
resolve_selection
ChatGPTParse one @aip reference line the user copied off a project page into data. A line names its project, optionally a moment in composition time, and either an element address or a frame region the user drew. references carries one entry per line in the paste, each with kind and the full address for that kind; composition_id is set only for a moment. A library card line has no time; read a motion-asset with get_motion_asset, a project-asset with get_project_asset. Refused when the line names a project other than project_id, or carries no readable reference. Nothing changes. Billing: free.
restore_version
ChatGPTReplace the live editable preview with one saved version, through the editor's own restore path. It takes the project lock and returns task_id to poll with get_task; while another task holds the lock it refuses with that active task id. Only a version_id from list_versions is accepted; any other value answers 404. Billing: free.
save_editing_preset
ChatGPTStores the calling account's editing preset. The object sent replaces the saved one whole, so a key left out of it is a key dropped; null, or an object with no answers in it, clears the preset. Returns editing_preset as it now stands. The preset is a record of what the user chose; it admits no stage, changes no price, and no other tool on this server reads it. caption_style=no-caption means skip caption generation and leave the caption layer unmounted, not a pattern. Billing: free.
sign_upload
ChatGPTSign a local .mp4/.mov/.webm upload. Byte limit: up to 3584 MiB, or 2040 MiB for restricted uploads; duration is separate (start_transcribe_project). At 15 minutes these allow ~33 or ~19 Mbps total average: compress or trim larger files. Returns upload_url, required PUT headers, video_ref, upload_expires_at (null=no deadline). No bytes move until your shell PUTs them; repeat PUTs before expiry replace bytes (single_use=false). Resolve missing authorization for this file/destination before sending; reuse prior authorization; a host denial is not consent. Billing: free.
sign_workspace_upload
ChatGPTSign PUT targets into one project's workspace. No bytes move: your shell PUTs each file to its upload_url, again if needed until upload_expires_at (last write wins); content_type is bound into the signature. Nothing is visible until commit_workspace accepts it. Writable paths are render-engine/index.html, render-engine/compositions/, render-engine/public/, render-engine/styles/, render-engine/fonts/. Other paths, and types no commit accepts (no script, no .svg: only markup, stylesheets, JSON, fonts, images, video and audio), land in rejected with a code; the rest in uploads. Billing: free.
start_export
ChatGPTRender the project's current editable preview to a downloadable mp4. A render this call starts passes render-tree admission first: a preview that fails it is refused with the blocking issues and no export is created. Admission is capped at 8 per 5 minutes per caller plus a shared server cap; over it the reply is a 429 with no verdict. An export already running, or still fresh for this preview, is returned without admission, and says so in reuse_reason. status deferred means a live turn owns the project and starts the export when it ends. Every reply that started, reused or parked an export carries task_id, current_workspace_digest (the accepted workspace NOW, the same token list_workspace reports), staged_files_present and warnings, so a refused commit, an upload nobody committed and an unchanged workspace no longer read alike. Those fields describe the workspace, not the returned file: this call does not yet tell you which workspace a given mp4 rendered, and naming the one you want is not supported. current_workspace_digest is null when the workspace could not be read, which does not affect the export. The file comes from get_export_url. page_url opens the project in the product app, with its navigation and without the AIP chat, and is the durable link to give a person in the final reply. agent_page_url opens the editor view for embedded panels; it is not the final delivery link. Billing: free.
start_project_from_attachment
ChatGPTCreate a project from an attached video when the host cannot move bytes; the server fetches it. One per call; .mp4/.mov/.webm to 3584 MiB (2040 restricted); duration limits as start_transcribe_project. A stub with no file is refused with a fix. Returns project_id, task_id, agent_page_url, page_url; get_task reports preparation progress. Voice leveled, never isolated: a clean voice is the separately billed enhance_voice. Billing: free.
start_transcribe_project
ChatGPTIngest/transcribe uploaded video_ref into an editable draft; needs a credit wallet. Source limit: 15 minutes; staff 60. Fixed per project, longer refused before billing. Returns project_id, task_id, agent_page_url, page_url (the durable link here to give a person); get_task reports preparation progress. Voice leveled, never isolated: a clean voice is the separately billed enhance_voice. Billing: free.
wait_task
ChatGPTWait inside one call for the exact task a mutation on this surface returned, answering with get_task's fields once status is one of done | failed | cancelled or wait_seconds elapses; a status change alone does not answer. wait.woke_on names what ended the hold. A caller already waiting on 8 tasks is answered at once with woke_on hold_limit; a wait past 4 on one task ends the caller's oldest with woke_on superseded. Sends a notification every 15 seconds while held. A reply that is not terminal carries next_data.poll_args, this wait again. Billing: free.
build_captions
Claudecancel_task
Claudecommit_workspace
Claudeenhance_voice
Claudefind_broll
Claudefinish_project
Claudeframe_speaker
Claudeget_chat_history
Claudeget_editing_preset
Claudeget_export_url
Claudeget_motion_asset
Claudeget_pricing
Claudeget_project
Claudeget_project_asset
Claudeget_task
Claudeget_transcript
Claudeget_view_url
Claudeget_workspace_file
Claudelist_motion_assets
Claudelist_project_assets
Claudelist_projects
Claudelist_versions
Claudelist_workspace
Claudematerialize_motion_asset
Claudepresent_choices
Clauderecord_choices
Clauderesolve_selection
Clauderestore_version
Claudesave_editing_preset
Claudesign_upload
Claudesign_workspace_upload
Claudestart_export
Claudestart_project_from_attachment
Claudestart_transcribe_project
Claudewait_task
Claude
