Everygen

Add captions or subtitles to a video edit. Use it to add captions to clipping, Ryu and live-editor projects, after requested cuts and reframing; the captions stay editable. Read get_editing_guide topic captions for style ids and supported settings; users can also change the look in the live preview inspector. Every clip with speech is transcribed once and the words land where they play. Library music and muted clips are skipped. Calling again replaces these captions; use restyle_edit_captions for appearance only. Uses the editor caption meter: 1 credit per 25 minutes heard, charged only after captions save. A failed call costs nothing. Pass the revision just read.

caption_edit

Convert speech in an audio file to a selected Everygen voice. Pass a voiceId from list_voices. For local files, first upload audio with prepare_generation_input_upload using kind referenceAudio.

change_voice

Story workflow gate (Phase P): verifies every generation prompt is within the 2500-character hard limit (2300 target) BEFORE calling generate_image / generate_video — over-limit prompts are rejected with an error. Pass all master prompts at once. Free and instant.

check_prompt_budget

Story workflow gate (Phase S3): checks narration text for AI-narrator tells — zero contractions, uniform sentence lengths, missing punch/long lines, banned filler phrases. Run on the formatted VO text before generating the voiceover; fix errors and re-run until ok. Free and instant.

check_voice_cadence

Everygen Clone: remake a video with the user's own people, outfits, products or places while keeping its exact motion and timing. transfer recasts everything; swap changes only what is named. Before calling, you need the video, as a public social link or in Everygen, and, unless the user only described a change, their photos in Everygen. For anything still missing, call upload_media_widget right away, for the video and photos together; never ask them to attach files to the chat: those can't be used. When they want their photo used, pass its media URL in images; without it, open upload_media_widget again instead of starting. Don't ask them to choose transfer or swap: use swap when they name one thing to change, transfer when they want the whole video recast. When it is unclear what a photo replaces, ask first with these choices: the main person, a product or object, the environment, or the environment and the person. The card above runs the whole Clone: it makes swapped frames from the video and shows them, and nothing renders until the user approves them there or tells you to. A swap Clone can edit directly renders straight away with no frames, so never promise frames before the card shows them. Charges the whole Clone when it starts and refunds it if the frames are never approved. Do not poll and do not describe the card; after calling, say one short line at most.

clone_video

Verify an uploaded file's ownership, media type and stored size. Large videos return pending while optimization runs; repeat this call with the same key until ready. Only the verified optimized video is saved and returned, and the temporary original is removed. A ready result includes an Everygen asset, URL and mediaInput for editing or clipping. Generation inputs must still meet each model's reference limits.

complete_generation_input_upload

Render a standalone final or a generated workflow's stitch. For editable combinations use open_edit with sources in play order and open_edit_preview, then export_edit when requested. In this standalone flow, give an ORDERED list of clips; each may carry audioUrl (a voiceover) which overlays that clip and mutes its own sound. Story workflow finals use TWO passes: pass 1 glues the masters (one clip per master, window = its locked plan duration, NO audioUrl on any clip); pass 2 is a SINGLE clip { videoUrl: the glued video, audioUrl: the one voiceover, durationSeconds: the plan's target total } which overlays the narration and mutes everything else. Never put audioUrl on a multi-clip payload for a story final — the audio would end with that clip. Also use it to put a fresh voiceover over an uploaded or de-captioned video (a single clip with an audioUrl). (Legacy plan_story_scenes conversations only: one clip per scene with its per-scene voiceover.) All URLs must be Everygen URLs from earlier steps (generate_video final URLs, generate_voiceover audioUrl, uploads, or caption-removal results). Max 3 minutes total. For story videos, do NOT call this until the user has reviewed the master clips and explicitly approved the final render. Captions on an explicitly requested standalone final use add_captions; editable captions use caption_edit. Optional overlays are on-screen text (rank, timestamp, or label). Optional imageOverlays place an https image as a sticker. Optional audioUrl is one soundtrack or voiceover for the WHOLE video and mutes every clip; volume (0–100) sets that track's loudness and is only valid with audioUrl. Omit overlays, imageOverlays, audioUrl, and volume and the call stitches clips exactly as before. Returns a self-polling video card.

compose_video

Save a reusable character from an image. IMPORTANT: to create a character you must FIRST have an image — call generate_image (or upload_media) to produce one, show it to the user, then call create_character with that image URL. The result renders the saved character image for confirmation. If description is omitted it's auto-generated (appearance only). A public image URL can also be fetched and saved.

create_character

Cut one moment from a completed find_clips job into a standalone video clip (up to 10 minutes), cut from the original source at full quality (not the low-res preview). Cuts up to 3 minutes are frame-accurate; longer custom ranges snap to the nearest keyframe. Pass the jobId plus either momentIndex (from show_clips, 0-based) or a custom startSeconds AND endSeconds pair — a single custom bound combined with momentIndex is rejected as ambiguous. Use this for a requested custom source range; selected moments that are already cut go straight to the editor. Returns the cut clip as a video. A long source can take a few minutes: if it returns status 'processing', wait about a minute and call cut_clip again with exactly the same arguments to get the finished clip. For captioning or editable refinement, call get_workflow with slug 'ryu-workflow' and phase 'start', then follow its MCP editing instructions. Reuse the matching live project when continuing the same clip; otherwise open_edit with the selected owned clip refs, then open_edit_preview once. Read the current revision with read_edit before timeline tools; use caption_edit to add captions and restyle_edit_captions for appearance changes. Saved edits update that same preview. Several selected finished clips get independent projects unless the user explicitly asks to combine them. Use export_edit only when the user asks for a downloadable final; the live preview needs no rendered export. Reserve place_clip/add_captions for an explicit standalone one-shot captioned-file request; select_caption_style is optional browsing within that path.

cut_clip

Delete one of the user's saved characters by id. Ownership-checked; only the caller's own characters can be deleted.

delete_character

Change a Clone's frames while they wait for the user, before anything renders. Pass the change as the user put it, saying only what changes and naming whoever it changes by what they wear or look like, since the image model sees only the frame. Edit every frame that shows the change, or the video mixes both versions: pass frames by name only when the user pointed at specific ones, otherwise leave frames out to edit them all (up to 12 per call). The card above shows the edits in place, so don't present them; ask whether they look right. Each frame is charged as one image edit. A wrong person replaced is not a frame edit; see get_clone_status.

edit_clone_frames

Open or switch the live canvas preview to the edit being worked on. Immediately after open_edit succeeds, call this with its exact projectId. When resuming a different saved edit, read_edit then call this once with that projectId. The agent chooses the project; there is no project picker. Keep this view open: later saved revisions appear automatically. Do not call again for each revision or render/export just to refresh it. Omit projectId only for an empty app entrypoint. Playback is free.

open_edit_preview

Edit the current project using the revision just read. Patch fields directly, e.g. /elements/2/rotation. Times are frames at 60 fps; keyframe time is relative to the element start. x/y are pixels from canvas center, y is up; opacity is 0-100. Use trim_edit for trims, never start/duration patches. Add media with add_to_edit. Removing a main-track clip closes its gap; captions follow cuts and moves automatically. reorder is one-based clip numbers in play order, e.g. [2,1,3]; send it alone with ops empty. transitions joins adjacent clips by element index or artifact:// ref. Change zoom, framing, text or effects here instead of generating a new video. get_editing_guide provides valid text defaults, catalogs, framing and keyframe examples. Writes must be sequential. Code scenes: copy to reuse (new id), edit-text {path,find,replace} to revise; never retype source. Check visual edits with view_edit_frame; the open live preview updates without a render.

edit_timeline

Read the editing workflow and native editor reference used for Supercomputer-style edits. overview covers tool choices, timing and revision rules; framing covers crops/resize/keyframes; text includes a valid text element; effects, transitions, stickers and captions expose the real catalogs and settings; audio covers music/SFX; code-scenes explains authoring HTML/CSS/JS animations directly with the host model. Read the relevant topic before an unfamiliar edit instead of guessing fields or ids. Free; never changes a project.

get_editing_guide

Estimate Everygen credit cost for image or video generation. Video model defaults match generate_video (Kling 3.0 Turbo when model is omitted, including when references are present). Only call when the user explicitly asks about cost or pricing.

estimate_generation_cost

Export the saved revision at full quality. TURBO costs 2 credits and HD 3; some accounts have one full-quality export for 1 credit instead, and the card shows what was charged. Charged when a render slot is claimed and refunded if the render fails. Returns an artifact id and a video card; poll get_edit_render with the operationId.

export_edit

Find the best moments in a long video AND automatically cut each one into a finished, ready-to-share clip (podcast, interview, stream — anything with talking). Give it a video URL (YouTube, Instagram, TikTok, X/Twitter, Facebook) or a Everygen-hosted video URL (from upload_media or a prior generation). Supports videos up to 8 hours, though it works most reliably under about an hour — very long social videos can be slow to fetch and may time out. Works best on speech-heavy videos; videos without talking will be rejected. Optionally pass `instructions` to look for a SPECIFIC kind of moment in plain language (e.g. 'every time someone gets voted out', 'product close-ups', 'the biggest laughs') instead of general best-moments — clips are then picked for relevance and each carries a matchReason. When the result includes `matched: false`, the instruction matched nothing and the clips are a generic fallback — TELL THE USER their instruction didn't match before showing the clips. For captioning or editable refinement, call get_workflow with slug 'ryu-workflow' and phase 'start', then follow its MCP editing instructions. Reuse the matching live project when continuing the same clip; otherwise open_edit with the selected owned clip refs, then open_edit_preview once. Read the current revision with read_edit before timeline tools; use caption_edit to add captions and restyle_edit_captions for appearance changes. Saved edits update that same preview. Several selected finished clips get independent projects unless the user explicitly asks to combine them. Use export_edit only when the user asks for a downloadable final; the live preview needs no rendered export. Reserve place_clip/add_captions for an explicit standalone one-shot captioned-file request; select_caption_style is optional browsing within that path. This returns a live widget that shows progress and then displays the finished clips automatically as soon as they are ready. If the Clip Moments widget is not on screen, call wait_for_clips with the jobId, then call show_clips once it returns status complete. Do NOT call cut_clip for these — they are already cut. The user picks which clips to edit or download from the widget.

find_clips

Search the editor's music and sound-effect library. kind is music (default) or sfx. query matches title and mood. Returns up to 20 tracks. Pass a track id to add_music_to_edit. Free.

find_music

For an explicit standalone one-shot captioned-file request, open a framing widget so the user positions and zooms a HORIZONTAL clip inside the vertical 9:16 caption frame before add_captions. Skip this when the standalone clip is already vertical/square. The widget sends back the add_captions call with the chosen placement. For ordinary editing or an existing live project, use the Ryu workflow and the project's timeline framing tools instead; do not replace the editable preview with this standalone captioning path.

place_clip

Generate a marketing video ad using the selected product, avatar, voice, format, and script. CRITICAL: the script must be SHORT enough to fit the selected duration - about 2-3 sentences for 8 seconds, 3-4 for 10 seconds. Do NOT write a long monologue. The character must finish speaking before the video ends. This tool builds the full marketing prompt with format-specific templates and role-tagged reference media. One call makes ONE continuous clip, up to 15s (some models allow less; see duration) - for longer ads make one call per scene with the same avatar and voiceId (see the long-ad workflow). firstFrameUrl pins the opening frame (whenever other references are attached - which is any ad with a product or avatar - it is applied as a locked reference image with a first-frame lock in the prompt, since providers reject a hard first frame combined with references). referenceImageUrls/referenceVideos/referenceAudios attach extra references like the in-app studio composer; video and audio reference clips are supported on the models listed on those parameters and require durationSeconds (use the values returned by the upload widget). Returns a video result card that self-polls. COST: this call starts ad generation and consumes credits when accepted. The confirmed parameter is currently ignored; do not call once just to request an estimate.

generate_ad

Generate a custom avatar image from a text prompt, or save an uploaded image as an avatar. Called by the ad builder widget.

generate_ad_avatar

Generate music with ElevenLabs Music or Lyria, or a sound scene with Seed Audio. For a song, soundtrack, or jingle pass model elevenlabs-music with durationSeconds; it takes no references. Pass lyria-3.5 (full song) or lyria-3-clip (30-second clip) only when the user asks for Lyria or wants up to 10 imageUrl references to set the mood. Custom lyrics belong in the prompt with [Verse]/[Chorus] tags. Omit model for ambience, sound effects, and multi-role dialogue (Seed Audio: up to 3 audioUrl clips or one imageUrl, never both). NOT for plain narration voiceovers (use generate_voiceover). Returns a stable R2 URL and renders an inline audio player. Lyria also returns the lyrics it wrote.

generate_audio

Generate a still image with Everygen. Use Nano Banana 2 by default, GPT Image 2.5 Flare for fast OpenAI generations, GPT Image 2.5 Sunburst when editing precision matters, Nano Banana Pro for highest-quality photorealism, and Seedream when the user asks for it (Pro for precise edits, Lite for 2K-4K and web search, Flash for the fastest render). Pass useSearch:true with a Gemini model or Seedream 5 Lite. Use this only for explicit still-image requests. If the user may want motion, use generate_video instead.

generate_image

Generate 2-8 independent images in one call and return every result in input order. Use this for clean plates, variants, storyboards, placement frames, or any job with multiple prompts. Each item can choose its own model, aspect ratio, resolution, search setting, and HTTPS reference images.

generate_image_batch

Generate a voiceover script using AI. Returns a written script optimized for short-form video (TikTok, Reels, Shorts). Provide a topic, product description, or paste a YouTube URL to analyze and write a script based on the video. Optionally pass a format and/or a channelStylePrompt (from analyze_channel_style) to match a specific channel's voice. For multi-scene story videos use the story formats (animal-story, bodycam, custom-story) — they apply the viral story formula where each paragraph maps to one video scene. The script can then be used with generate_voiceover to create audio.

generate_script

MODEL SELECTION: Keep a model already named by the user, a workflow, analyze_video, or a story plan. Do not replace that model with Kling 3.0 Turbo or MiniMax. Only when nothing named a model, pass Kling 3.0 Turbo, including when reference media is attached. Do not switch that omitted model to MiniMax just because references are present. Kling 3.0 Turbo accepts first frames but not reference images; for a reference sheet pass Kling 3.0, MiniMax H3 Max, or Seedance 2.5 explicitly. Reuse the chosen model for subsequent scenes. Always pass the model explicitly. Do not include captions or subtitle wording in video-generation prompts unless the user explicitly asks the video model to generate them. Start an Everygen video generation and return generation ids for polling. Keep a model already named by the user, a workflow, analyze_video, or a story plan. Only when nothing named a model, pass Kling 3.0 Turbo, including when reference media is attached. Do not replace a named model, and do not switch the omitted-model fallback to MiniMax because references are present. Kling 3.0 Turbo accepts first frames but not reference images; for a reference sheet pass Kling 3.0, MiniMax H3 Max, or Seedance 2.5. For ads/product videos use generate_ad instead. IMPORTANT: pass orientation="portrait" when the user asks for 9:16, and orientation="landscape" when they ask for 16:9. Seedance and MiniMax H3 Max accept ratio strings. For story videos, follow the story workflow (get_story_workflow phase "start") — it locks camera medium, continuity, and durations before any generate_video call, and every prompt must pass check_prompt_budget first. LEGACY: scenes planned by plan_story_scenes pass storyPlanId + sceneIndex instead of a prompt (the server injects the exact planned prompt; those scenes default to Kling 3.0) — only to finish conversations that already hold a planId. COST: this call starts generation and consumes credits when accepted. The confirmed parameter is currently ignored; do not call once just to request an estimate. If later workflow steps need this video, keep the turn active and repeat bounded wait_for_generation calls on its returned IDs until terminal, then continue. Do not depend on a card notification to resume. Veo models need duration 8 when lastFrameUrl is set. Unsupported durations are snapped; when that happens the result includes duration (what ran). Report that length.

generate_video

MODEL SELECTION: Keep a model already named by the user, a workflow, analyze_video, or a story plan. Do not replace that model with Kling 3.0 Turbo or MiniMax. Only when nothing named a model, pass Kling 3.0 Turbo, including when reference media is attached. Do not switch that omitted model to MiniMax just because references are present. Kling 3.0 Turbo accepts first frames but not reference images; for a reference sheet pass Kling 3.0, MiniMax H3 Max, or Seedance 2.5 explicitly. Reuse the chosen model for subsequent scenes. Always pass the model explicitly. Do not include captions or subtitle wording in video-generation prompts unless the user explicitly asks the video model to generate them. Start 2-8 distinct video clips in one reliable server-side batch and return every accepted generation id in scene order. Use this instead of making multiple parallel generate_video calls for explainers, stories, montages, or any composition with multiple scene prompts. One scene object equals one clip; do not claim a scene was submitted unless it appears in sceneGenerations. When the workflow needs these clips for another step, keep the turn active and repeat bounded wait_for_generation calls for all accepted IDs until every clip settles, then continue. Do not depend on the card staying open or sending a wake-up message. COST: this call starts all accepted clip generations and consumes credits for each. The confirmed parameter is currently ignored; do not call once just to request an estimate.

generate_video_batch

Generate a WAV voiceover audio file from text using an Everygen voice and return a stable R2 URL. The result includes a timing field (speechEndSeconds, tailSilenceSeconds, sentenceEnds, beatEnds) — the story pack fits each shot to timing.sentenceEnds and durationSeconds. durationSeconds includes trailing silence; timing.speechEndSeconds is where speech actually ends.

generate_voiceover

Check a caption removal job and return its status and result video URL. This call may finalize a completed job or settle and refund a failed or stale job.

get_caption_removal_status

Check a standalone captioned-video render and return progress and its final video URL. This call may copy a completed video into the user's storage or refund and settle a failed render.

get_caption_render_status

Check the status of a find_clips job. Returns { status: 'processing' | 'complete' | 'failed' } (processing includes a stage: downloading / transcribing / picking / cutting). Poll this every 10-15 seconds until status is 'complete' or 'failed'. This returns plain data with NO UI — do NOT describe a widget. When status becomes 'complete', call show_clips ONCE to display the finished clips to the user. For instruction jobs, the complete result includes `matched`; if it is false, tell the user their request matched nothing (the clips shown are a generic fallback) before show_clips. If 'failed', tell the user the error. Never call cut_clip for these clips — they are already cut.

get_clips_status

Where a Clone is: preparing, frames_ready (with each frame's name and time in the video), rendering, completed, stopped or failed, and what to do next. The Clone card tracks this on its own, so call it only when the user asks about a Clone or before editing its frames. Never paste the URLs it returns.

get_clone_status

Check a compose_video render and return progress and its final video URL. This call may refund credits if the render has failed.

get_compose_status

Return the authenticated Everygen account's current credit balance. Only call when the user explicitly asks about their credits or balance.

get_credit_balance

Check a preview_edit or export_edit render. This call may copy a finished video into the account or refund a failed render. Pass the operationId from the card.

get_edit_render

Compatibility alias of get_workflow with slug "gta-6" (the GTA 6 house pack). Prefer get_workflow. A phase that pack does not have returns its start phase. Free — reading instructions never charges credits.

get_gta6_video_workflow

Fetch status and final URLs for Everygen image generation ids. A timed-out generation may be marked failed and refunded during this check. Does not render a card - the generate_image result card updates itself.

get_image_generation_status

Compatibility alias of get_workflow with slug "nursery-rhymes" (the nursery-rhymes house pack). Prefer get_workflow. A phase that pack does not have returns its start phase. Phase "song" returns that house-pack phase. Free — reading instructions never charges credits.

get_nursery_rhyme_workflow

Compatibility alias of get_workflow with slug "story" (the found-footage story house pack). Prefer get_workflow. Call with phase "start" first, then the phase that step names (script, look, shots). An older phase name (story, storyboard, continuity, model, plates, realism, prompts, delivery, failures, contracts) returns start. Free — reading instructions never charges credits. On ChatGPT, 2–8 shots are one generate_video_batch.

get_story_workflow

Check a video analysis job started by analyze_video. Read generations[0].status: pending means keep polling every 5-10 seconds; failed is terminal and includes an error; completed includes the full analysis at the top level. Read recreationMethod and recreationGuide before choosing a recreation workflow. A completed result includes script: when the user asked for the script or transcript, paste script in full. Do not treat pending as failure. Recreation confirmation: after analysis, show the exact proposed spoken dialogue/narration in scene order, separately from visual directions (or explicitly state there is no speech). For a faithful recreation, preserve the source wording by default; for an adaptation, label the rewritten lines. Flag uncertain transcript words instead of inventing them. Summarize the planned visuals, avatar/voice, total duration, and any changes from the source, including omissions or a shorter sample. Ask the user to approve this concrete script and scope, then STOP and WAIT before generating images, voiceovers, audio, or video. A request to recreate, an avatar/voice selection, or cost approval alone is not script approval. Never silently replace the dialogue or reduce a full recreation to a test clip. Reuse explicit approval of the displayed script and scope; ask again only if either changes. Analysis-only requests need no generation approval.

get_video_analysis_status