Cadabra AI

Use this when the clips already exist and need cutting into one film. Not for re-editing a finished MP4 (cadabra_edit_film). Cut already-made clips into one film: dissolves, captions, music ducking. Same editor as cadabra_make_film. A link to the creator's OWN site, app or YouTube video written in the brief is opened on Cadabra's render server to take phone-sized screenshots and read the page text for the cut (a YouTube link gives only its title and thumbnails); nothing else is visited and nothing is downloaded from the video. A batch link is fine. PHOTOS WORK TOO (jpg, png, webp): each one is held and moved slowly on screen, so a set of photos becomes a film for an event, a memorial, a birthday or a property listing (recipe photo-film). Always pass aspect_ratio with photos. Price is the existing edit fee: 5 credits base, +1 per extra 15 seconds of finished runtime. Set confirm:true to spend. Without it you get a quote only.

cadabra_assemble

FREE. The two pieces of writing a business needs before any post exists. · brand — what this brand SOUNDS like: its voice, what it promises, who it is talking to. Run it once and reuse it. · ads — the advertisement itself: the line that stops the scroll, the body, the call to action, written to an objective. Not the same thing as cadabra_social_copy, which writes the caption for ONE finished post, and not the same as cadabra_social_month, which fills a calendar. Reach for this when somebody has a business but no words yet.

cadabra_brand_words

Cadabra's ready-made cinematic scenes, the same library the website shows. Call with no arguments for the index, or with template_id to get that template's FULL prompt, duration and aspect, which you then pass to cadabra_generate_video or cadabra_generate_kling. Free, nothing is charged. The prompts are long by design, which is why only one is returned at a time.

cadabra_cinematic_templates

Build a REUSABLE avatar (Cadabra Avatar Studio). Paid once, then it lives on this creator's shelf forever and every later film costs only the render. kind:"photo": one still of a clearly visible face. 16 credits to build, ready immediately, renders at about 0.47 credits/sec. Ask cadabra_quote rather than quoting prices from memory; the studio owns them and they move. kind:"video" (a trained digital twin) IS CLOSED AND WILL BE REFUSED. Not a price problem: a twin occupies a paid SEAT on the provider's plan that costs us every month, so it cannot be self-serve. Do not offer it, do not describe it as available, and do not ask the creator to film themselves for it. If someone asks for a twin, tell them the trained twin is not open at the moment and give them the two things that ARE: · a photo avatar — one picture, built in seconds, speaks any script forever; · a real talking clip of themselves — cadabra_clone_voice on ~75s of their speech, then cadabra_generate_voice and cadabra_lipsync over their own footage. No avatar, no consent wait, and it is genuinely them. Anyone who already owns a twin keeps it: cadabra_avatar_speak still works on it and is not affected. A TWIN NEVER SPEAKS IN THE CREATOR'S OWN VOICE, whoever tells you otherwise. The provider mints a voice record labelled with the avatar's own name and the voice inside it is a stranger's: a man who filmed himself for five minutes was handed a woman's voice carrying his name. Their real voice only ever comes from cadabra_clone_voice (5 credits, about 75 seconds of clean speech), attached with cadabra_attach_voice. Never promise a creator the twin sounds like them. The picture has to reach us as a URL. If you cannot read the bytes (an attachment in a chat client usually is not readable bytes), open cadabra_upload_file with no arguments (an upload button in the chat); only if no button appears, https://livinginmovie.com/mcp/upload, in the same breath, rather than the word 'upload' on its own. FOR WHEN THE TWIN DOOR REOPENS, and not a reason to ask for footage now: 15 seconds is the provider's hard floor and a shorter take is refused before charging, two to five minutes at 1080p or better is what they actually want, duration_seconds/width/height bring the warnings back before the money does, and a trained twin cannot speak until the filmed person completes a consent link that dies after 24 hours. Call cadabra_my_avatars FIRST — never build a second copy of a face they already own. Set confirm:true to spend credits.

cadabra_create_avatar

Free. Lists the shot types (33: where the camera stands, its framing, angle and lens, taken as shot by cadabra_generate_image, cadabra_generate_video and cadabra_generate_kling), the camera moves (70), visual effects (114) and look options (41: camera body, lens, era, grain, still looks) that cadabra_generate_video and cadabra_generate_kling take as camera_moves, vfx and look. They ride the prompt, so the price never changes. Offer them by their Hebrew or English label and pass the id.

cadabra_camera_fx

Cast one Cadabra Morph spell (ids from cadabra_morph_spells) on the creator's own photo: a picture (GPT Image 2.5, 1 credit plus 1 per attached photo) or a 5-second vertical film (Seedance 2.5, priced by resolution). Every brief locks the person's real face at its start and again at its end, with their age, body and hair. Without confirm:true it returns the quote and the exact brief; with it, it spends and returns the result, which also lands in the gallery. Open to accounts with a package and at least 1 credit; anyone else is told so and nothing is charged. Use only photos of the creator themself, their pet, or someone who agreed to it.

cadabra_morph

Free. Run this on any text that is about to be read aloud by cadabra_generate_voice or cadabra_generate_dialogue. A speech engine voices what is written, letter for letter: a typo becomes a mispronounced word, a number becomes the wrong number out loud, and the mistake is only heard after the credits are spent and the MP3 is in the creator's hands. This returns the corrected text, a short list of what was changed, and — when the writing is chatty rather than spoken — an alternative written in narration register. Up to 5000 characters. Hebrew included.

cadabra_proofread_script

Free. Hand it a frame from a shot you just rendered (cadabra_inspect gives you frames) and, when there is one, the reference photo the person was supposed to match. It answers with the one thing that is wrong, or with nothing. Pass an end frame too whenever you can. The two failures that actually reach a creator — the character changing halfway through, a camera move that never resolves — happen ACROSS a clip and are invisible in a single frame. It is advisory and it fails open: silence means no problem was found, never a guarantee. It does not block anything and it never re-renders on its own. Use it to catch the shot before the creator does, then tell them what you found and let them decide.

cadabra_qa_shot

Use this when they ask how many credits are left, and after any failure, before saying anything about money. Return the signed-in Cadabra user's remaining credits.

cadabra_check_balance

Turn a recording of a real voice into a voice you can then speak any text with. 5 credits. Only clone a voice the user owns or has permission to use. Thirty seconds of clean speech beats five minutes of noisy. Set confirm:true to spend credits. HOW THE RECORDING GETS HERE, because this is where people get stuck: you need a URL. If you can read the file's bytes, cadabra_upload_file makes one. YOU USUALLY CANNOT (audio attached in a chat client is not readable bytes), so open cadabra_upload_file with no arguments (an upload button in the chat); only if no button appears, https://livinginmovie.com/mcp/upload (same account, drop the recording, paste the link back), straight away. Never reply 'upload the recording' without it.

cadabra_clone_voice

The final cut of a music clip, built in code, not guessed. Before planning ANY music clip, read cadabra_editor_manual section 'music-clip' (free): how international-level clips are made and how to lead the creator there. Every every singing shot starts 0.3 s before its line (the slice it was made from starts there too) so the mouth lands on the words; the gaps are filled with B-roll cut to whole bars at the song's own tempo; every clip is muted and the song carries the sound, mastered loud for social. Pass range_start/range_end for a vertical 15 to 30 s cut from the best part (usually the chorus). Same edit fee as cadabra_assemble. Without confirm you get the quote, the runtime and the shot order.

cadabra_music_video_cut

Cadabra Morph Mix: one clip the creator filmed, restyled in 2 to 4 styles (style ids from cadabra_video_effect's style list), then cut between them on every snap or clap with the original voice underneath, as one reel. Each style is priced like a restyle of the whole clip on Seedance 2.5; the joining is free. Without confirm:true it returns the quote only. Takes minutes; the result also lands in the gallery, and each style on its own as well. Open to accounts with a package and at least 1 credit.

cadabra_morph_mix

Permanently remove ONE item from this creator's own gallery. Free. THERE IS NO UNDO AND NO TRASH. The row is gone. So: take the id from cadabra_list_creations, tell the creator exactly WHICH item you are about to delete in words they can check (what it is, when it was made), and only send confirm:true after they have said yes to that specific thing. Never work from a description. "Delete the flower one" is not something anybody should be able to say to an irreversible action; find it, name it, confirm it, then delete it. It can only ever touch this creator's own gallery; the scoping is the studio's, not this tool's.

cadabra_delete_creation

Turn a photograph into a per-pixel depth map. 1 credit. What it is for: convincing fake depth of field, a parallax move on a flat photograph, and as a mask that knows what is in front of what. Returns an image where near is bright and far is dark.

cadabra_depth_map

Invent a voice for cadabra_gemini_speak from a written description and get a preview to listen to. It costs credits (Sela 2026-09-28: design is never free): call without confirm for the quote, show the price, and confirm:true only after the creator's yes. The preview lands in their gallery; a failure is refunded. Describe it in English: age, the language and accent it speaks natively ("native Israeli Hebrew, a Tel Aviv accent"), pitch, pace, the feeling. For a Hebrew script this is the proven route: a voice designed as a native Hebrew speaker. Play the preview to the creator before using the voice. Hand back the voice_... id and keep it in the conversation: it is not in cadabra_gemini_voices and nothing else remembers it for them.

cadabra_gemini_design_voice

Design a scroll-stopping cover image for a video or post. Cadabra's thumbnail director writes the concept (free), then each concept is rendered as an image. 1 credit per concept, 2 when a reference photo is attached. Returns the artwork plus the headline text written for it. Note: on the website the headline is burned onto the picture by the browser; here you get the image and the words separately, so you can place them yourself. Set confirm:true to spend credits.

cadabra_make_thumbnail

Free. The rooms that come BEFORE anything is generated, where the thinking is cheap and the mistakes cost nothing. mode "document": a REAL document becomes a film you can shoot. Paste a treatment, a client deck, a chapter, a research paper, a proposal (up to 120,000 characters, so a whole chapter fits) and it comes back mapped: what the document actually says, the evidence and VERBATIM quotes it contains, what it deliberately does NOT cover, plus a world and a cast built to carry it. This is the one to use when the creator arrives with material rather than an idea. Pair it with cadabra_read_pdf when the material is a PDF. mode "bible": a treatment or story becomes a compact project bible: logline, one style lock in English, up to four characters and four locations described by LOOK. This is the cheapest way to hold a face and a place steady across a whole film, and it belongs before shot one, not after the second shot drifts. mode "script": one line becomes a shootable 30-60 second script, 3 to 5 shots, one named character reused in every shot. mode "research": a sentence, ideally with a URL or two in it, comes back with what is actually out there. It READS the links you paste. Use it before writing an ad in a category the creator knows better than you do. mode "expert": one of five departments answers on the brief in their own discipline: strategist, screenwriter, cinematographer, director, animator. Ask when a brief is thin and you need the question behind it, not more adjectives. mode "concepts": THREE NAMED DIRECTIONS from the creator's own words, before any detail is asked. Five ideas are written and a cold critic keeps the best three, ranked, each with a title, a one line logline, the opening (the first three seconds), the turn, and a score; the first is the recommended one with its reason. Show all three, recommend the first, and let the creator pick or keep their own. What they pick is CARRIED FORWARD, never decoration: it answers the turn (unless they already gave one), its opening becomes the first beat, and on a trailer its title becomes the title card while its turn is WITHHELD (a trailer promises, it never reveals). Then draft the beats of the form, let them rewrite any line, and only then price. All of it is text. Nothing is rendered and nothing is charged.

cadabra_develop

Generate a new LOOK for an avatar the creator already owns: the same face, in a different outfit, a different room, a different framing — and it can speak, like any other look. THIS IS WHAT MAKES A FILM INSTEAD OF A TALKING HEAD. A trained twin otherwise wears whatever it was filmed in forever, so any script that needs the person in pyjamas and then in a studio looked impossible. Generate one look per situation, then pass each look_id to cadabra_avatar_speak for that segment. The identity is carried by the avatar's existing look used as reference, not by your words — so describe the CLOTHES, the PLACE, the LIGHT and the FRAMING, and never re-describe the face. About 13 credits per look. Set confirm:true to spend credits.

cadabra_avatar_look

The six edit modes of the website's Midjourney V8.1 studio. 1 credit each, quote first, auto-refunded if it fails. - variation / remix: a new take on ONE of the four images of a Midjourney grid made earlier. Needs source_task_id (the taskId cadabra_generate_image returned for that midjourney run) and image_number 0-3 (top-left 0, top-right 1, bottom-left 2, bottom-right 3). strength "subtle" stays close, "strong" (default) moves further. remix also takes a new prompt. - edit (canvas): repaint one grid image from a prompt. Needs source_task_id, image_number and prompt. - retexture: keep the shapes of ANY picture and restyle its surfaces from the prompt. Needs image_url and prompt. - remove-bg: cut the subject out of ANY picture. Needs image_url. - upload-paint (mask): repaint part of ANY picture. Needs image_url and prompt; mask_url is an image whose opaque area is repainted (without it the whole frame is repainted). - version "8.2": for a picture made with model midjourney-v8.2 (its grid only accepts V8.2 edits). Default 8.1. A V8.2 variation, remix or canvas edit is drawn by Midjourney V7, at the same price. Say the price (1 credit) before you call with confirm:true.

cadabra_midjourney_edit

Several characters talking in ONE audio file on Eleven v4: each line in its own voice, the feeling carried from one speaker to the next, directed by tags inside each line. Hebrew supported, up to 40 lines and 10,000 characters. cadabra_generate_dialogue is the eleven_v3 version and stays as it is. To redo ONE line, call cadabra_eleven_v4_speak with that line and that speaker's voice: v4 keeps it the same person. Pass voice_name per line, never ask for ids. Priced at 1 credit per 750 characters across all lines. Set confirm:true to spend credits.

cadabra_eleven_v4_scene

The Eleven v4 speech model as its own tool. 90+ languages, Hebrew included, up to 10,000 characters in one read. It performs the text the way an actor would, directed by tags written INSIDE the text, and it keeps the speaker the same person across every take, so asking for another take of a line never drifts to a different voice. Use it when the creator asks for v4, for the most acted, emotional read, or for a long narration that has to stay steady. cadabra_generate_voice is the eleven_v3 tool and stays the default; this one never replaces it. A new take of ONE line of a longer script: call this tool with just that line and the same voice. NEVER ask the user for a voice id. Pass voice_name; a wrong or ambiguous name returns the candidates, free. Any voice speaks any of the 90+ languages: the creator picks the voice, and it is never refused for the language of the text. Priced at 1 credit per 750 characters, the same as v3. Set confirm:true to spend credits.

cadabra_eleven_v4_speak

FREE. The complete manual of Cadabra's render engine, the same one our edit brain works from. Start with section "overview" (a map of every capability: picture, transitions, titles, captions, MOTION DESIGN with graphics, charts, chat, mockups, logo reveals, 411 animated stickers, particles, light leaks, film looks, camera moves, sound, end cards). Then read "shape" for every EditPlan field and its allowed values before writing a plan for cadabra_render_plan. Other sections: doctrine (how to cut), story, rules, recipes (named shot recipes), craft (feature-film editorial craft), motion (motion-graphics principles), limits (what the engine cannot do; a key outside the shape fails the whole render).

cadabra_editor_manual

Turn a list of clips into a Premiere / Final Cut XML sequence and hand back a link to the file. FREE. This is how somebody finishes in their OWN editor. Until now a creator who wanted to take a Cadabra cut into Premiere had to rebuild the timeline by hand, clip by clip. Offer it whenever they mention their editor, a client handover, or wanting to keep working on it elsewhere. The clips are listed in order with their real lengths; the XML points at the same URLs, so their editor pulls the media straight from the gallery.

cadabra_export_xml

Takes a picture and gives it MORE PICTURE: the same photo at a different shape, with the new space painted in to match rather than cropped, letterboxed or stretched. 2 credits. This is the answer to the most common thing a creator hits: a portrait photo that has to run as a 16:9 cover, a square product shot that has to fill a 9:16 reel. Cropping throws away the subject, bars look like a mistake, stretching is worse. This keeps every original pixel and invents only the room around them. The subject is never scaled down to make room, and the original is never touched. Say where the picture should sit when it matters: a portrait becoming a landscape usually wants anchor "top", so the head stays high instead of floating in the middle with air above it.

cadabra_extend_image

Free. Full copy-paste prompts for stylised production-CG animation. kind: direct (story craft), dialogue (lock every line), master (full-body grey studio), sheet-a (three-panel headless), sheet-b (turnaround plus eight expressions), location (empty of people), prop, storyboard-sheet (eight painted panels), seedance-25 (named shots, 30s ceiling). Always returns the WHOLE prompt. Then generate with cadabra_generate_image look animation or cadabra_generate_video engine seedance-25. Quote before confirm.

cadabra_animation_craft

CLOSED since 2026-09-19: this tool always refuses and charges nothing. Do not offer it and do not quote it. It stays registered only so that a creator who asks for "my avatar filmed in a new scene" hears the route that does it: a trained character (cadabra_create_character once, then cadabra_generate_video naming that character), which also holds the same face across every later shot. For a cinematic feel on footage that already exists (a camera move, slow motion, a grade, a crash zoom), use cadabra_edit_film. Both quote their own price before anything is charged.

cadabra_avatar_cinematic

Use this when they ask for something they already made: the gallery, a past result, its link. It returns their real files: show those, never draw or illustrate pictures of your own in their place. Not for a job still running (cadabra_film_status). The user's gallery, searchable, with the PROMPT each thing was made from. Free. This is how you continue from past work instead of starting from a blank page: search for what they are referring to ('the coffee one', 'yesterday's reel'), take its URL as the input to the next tool, or take its prompt as the starting point for a variation. Whenever the user says 'that video', 'the last one', or 'like the one I made', come here first rather than asking them to paste a link. A film from cadabra_make_film is NOT here until it is finished. The gallery holds each shot as it lands; the assembled film is a JOB, and it appears here only once assembly completes. Never conclude from this list that a film failed — this listing names any film job it can see, and cadabra_film_status is the only truth about one. Telling a user their paid film 'was not created' while it is still assembling is the worst thing you can do with this tool.

cadabra_list_creations

Use this when someone has a LONG recording (a podcast, a talk, an interview, a lesson, a vlog) and wants the short clips worth posting from it. Not for a clip that is already short (cadabra_edit_film), and not for making new footage. It times every word (the same transcription cadabra_transcribe sells, at the same price, 1 credit per started 10 minutes), then finds the standalone moments for free: each one starts on a sentence and ends on one, keeps to the length asked, never overlaps another, and carries an honest score out of 100 in four named parts (hook, standalone, payoff, energy), a hook title in the recording's language, and why a stranger keeps watching. A flat recording is told it is flat. Without confirm it returns the price, free. Then show the moments with their scores and let the creator pick. Each moment comes with the exact cadabra_cut_reel call that turns it into a publishable vertical reel (the words are kept, so the reel never pays to hear them again); that reel is quoted by its own tool. For several reels at once, put those calls in a cadabra_mission.

cadabra_find_moments

Free. 44 finished looks for a picture of a REAL PERSON, each one a full production prompt with the wardrobe, lighting, lens and grade already written: the professional headshots (LinkedIn, CEO, lawyer, doctor, realtor, speaker, dating), the fashion ones (editorial, black and white, noir, golden hour, 70s, street), and the viral ones people ask for by name (action figure in the blister pack, animation still, brick minifigure, trading card, movie poster, magazine cover, superhero, cyberpunk, samurai, royal). Call with no arguments to see the shelf. Call with a look to get the prompt, then run it yourself on cadabra_generate_image with the person's photo in reference_image_urls and raw_prompt:true — the prompt is already crafted and the free craft pass would only rewrite it. Without a reference photo this makes a stranger in a nice suit, so ask for the photo before running anything. The prompts have male and female variants because wardrobe and styling differ. Use what you already know about the person from the conversation; ask only if you genuinely do not know.

cadabra_portrait_looks

Turn a rough thought into five DIFFERENT film ideas, each with the reason it would work and the reason it might not. Free, nothing is generated. Use this before cadabra_make_film whenever the user is not sure what they want: choosing between five costs nothing, and regretting a film costs credits.

cadabra_concepts

A flubbed word, a wrong number, a name said wrong: replace just those words inside a video or an audio recording without re-recording. The words are timed and the phrase is FOUND in them; the corrected words are spoken in the speaker's own voice (a saved voice by name, or cloned from this very recording) and laid into exactly that stretch, matched to the room's level and faded so there is no click. On a video the new words are fitted to the old stretch's length so everything after stays in sync, and with lipsync:true only a short window around it is lip-synced so the mouth says the new words. 🚨 ONLY on the creator's explicit ask, and only the words they name: an attached recording is theirs and is never rewritten on our own initiative. Without confirm you get an itemized quote (timing the words, a voice clone when no voice is given, the new speech, the lip sync); nothing is charged if the phrase is not found.

cadabra_fix_words

Generate a premium 4K video with Kling 3.0. Text-to-video or image-to-video, or Omni mode with up to 4 reference images (reference them in the prompt as <<<image_1>>>, <<<image_2>>>, etc.). Set confirm:true to spend credits.

cadabra_generate_kling_4k

Generate a video with Kling O3 (Kling Omni V3) — text-to-video, image-to-video, or reference-to-video (transform an existing video). Optionally reuse a trained character via element_ids (see cadabra_create_character). Set confirm:true to spend credits. Failed jobs are auto-refunded. 🎬 THE CODES RUN HERE TOO. A slash code (/fpvdrone /glambot /liquid /bullettime …) with mode image-to-video and their picture as image_start writes the whole English shot, compiled for Kling's own strength — physical simulation, weight, contact, cloth and liquid — rather than for a Seedance prompt. cadabra_video_codes lists all of them.

cadabra_generate_kling

Write a full song on Suno V6, the newest Suno. 1 credit flat, and flat really means flat: a 20-second jingle and a 6-minute track cost the same, so ask for the length the film actually needs instead of trimming to save money. Every run returns TWO different takes of the same brief for that one credit, so play both before deciding. Simple mode (custom_mode false): describe the song in a sentence and Suno writes the words and the music. Custom mode (custom_mode true): style and title are required, and your `lyrics` are sung as written, which is the only way to control the words. duration (10-360s) works in custom mode only, and V6 honours it. Suno sings Hebrew, unlike the video engines: Hebrew lyrics in custom mode come back sung in Hebrew. 🎬 A SONG WRITTEN FOR A VIDEO: send video_url in simple mode and Suno watches the clip and writes the words and the music to fit it (their limits: one clip, mp4/mov/webm, up to 100MB and 241 seconds). Say the language and the feel in prompt. Simple mode does not take a length, so lay it under the clip with cadabra_assemble and let the edit end it. Three models: V6 (best all round), V6_MINI (faster, lighter), V6_WILD (stranger, bolder). Afterwards cadabra_music_v6 extends it, covers it, splits the stems, makes a video of it or turns the singer into a reusable persona. Set confirm:true to spend the credit.

cadabra_generate_music_v6

Turn a photo + an audio clip into a talking-avatar video, in one call (Cadabra Avatar Studio). The photo becomes an avatar on the creator's shelf, paid once (16 credits), and it speaks the audio at about 0.47 credits a second. The same photo again skips the build, and the avatar keeps speaking new scripts through cadabra_avatar_speak. The photo needs a clearly visible face. Set confirm:true to spend credits.

cadabra_generate_avatar

Use this when they want ONE clip, or are working shot by shot themselves. A whole film, even on an engine they named, is cadabra_make_film with engine set. Not for changing a finished film (cadabra_edit_film). Generate a video. confirm:true to charge, else quote only. Failed jobs auto-refund. 🎬 A PICTURE + A CODE = A SHOT. If the creator wrote a slash code (/fpvdrone /tvcommercial /slowmotion /360 /dollyzoom /spiralshot /glambot /ugc /thirdpersonview /pov /gimbalshot /flyover /zoomout /orbitshot /macrozoom /macroreveal /earthzoom, or the Hebrew name of one), send it as `code` with their picture as image_url. The code writes the whole English prompt, locks the identity in that picture and picks the engine, length and frame. Their own words, if they wrote any, go in `prompt` and become the subject line. cadabra_video_codes lists all seventeen with what each one wants attached. Legal engine values: seedance-25 (Seedance 2.5), ltx-25 (LTX-2.5), seedance (Seedance 2.0 cube), seedance-4k, flux3, minimax-h3, gemini-omni, grok-imagine, happyhorse, wan-3, vidu-q4 (Vidu Q4, needs a picture). If they named 2.5 send seedance-25. If they named 2.0 or Seed 2 send seedance. 15s on seedance-25 bills as 20 then cut; do not switch to seedance to keep 15. Default engine seedance (2.0). Feature animation: get the prompt from cadabra_animation_craft kind seedance-25. One block equals one shot. Whole seconds, last timecode never above 30. Production CG look, no film stock. Dialogue only in braces. Attached files: their photo as image_file, a song or voice as audio_files, extra pictures or clips as reference_files. There is no separate upload step. They are copied to Cadabra's CDN on the quote call and their URLs come back in the quote, so the confirm call can send those URLs or the same files.

cadabra_generate_video

Worlds (Marble / World Labs): build a walkable 3D set from a prompt or a plate. Returns a camera still (not the odd 360 thumb). Video is NOT included. After this lands, call cadabra_generate_video with that still as image_url (engine ltx-25 or seedance-25). Two charges: 16 credits for the world, then the video engine. Set confirm:true to spend. Failed jobs are auto-refunded.

cadabra_generate_world

Use this when they want a still: a picture, a poster, a product shot, a character master or sheet. Not for moving pictures. Generate an image with Cadabra. Default model nano-banana (fast, cheap). look animation selects gpt-image-2 unless they named another model, and stamps production-CG laws (empty hands, no ink on skin). gpt-image-2 and both gpt-image-2.5 tiers charge +1 credit per reference image attached (on top of the base resolution/quality price) — tell the user this before attaching many references, e.g. a character sheet built from several photos. The 2.5 tiers sit just above gpt-image-2 for the same cell (2K high: flare 5, sunburst 8, against 4), so reach for sunburst when a real face, product or logo must survive an edit, when a mask_url marks the one area to change, or for dense small text, and say the credit number before you run it. Set confirm:true to actually spend credits — without it you get a price quote only. Failed generations are auto-refunded.

cadabra_generate_image

Use this only when they ask for the older V5.5 music model by name. A new song or score goes to cadabra_generate_music_v6, the default. Generate a full song from a prompt, or from your own lyrics (custom mode). Flat 1 credit. Set confirm:true to spend it.

cadabra_generate_music

Generate sound effects and ambient audio. Priced ~3 credits/minute of AUDIO PRODUCED (minimum 2), refined to the exact length after generation. A 5-minute worst case is reserved up front and refunded down, so the reservation is not the price — quote the estimate, not the hold. Set confirm:true to spend credits. 🚨 NOT THE NARRATION TOOL. This speaks with a generic preset voice. Any spoken line — narration, voiceover, dialogue, Hebrew or English — belongs in cadabra_generate_voice, which uses the account's OWN saved and cloned voices (call cadabra_list_voices first, it is free) and costs 1 credit per 750 characters, cheaper than this. A user asked for a specific saved voice by name, was brought here instead, saw the 15-credit reservation, and shipped a film with no voiceover at all. Come here for a door slam, a dragon wing, a room tone. Not for a voice.

cadabra_generate_sfx