agent_add_images
ChatGPTAdd the user's photos (public https URLs, at most 4; local files first go through create_image_upload_url) to the project as user_supplied assets.
agent_approve_storyboard_plan
ChatGPTRecords that the user approved a storyboard version (approval_id from request_storyboard_approval; when user_words names one, as "(approval_id=<id>)" or on older cards "(승인 <id>)", that one). user_words: the user's message about it as written, typed or sent with the card's button. Paid takes run against a recorded approval.
agent_cancel
ChatGPTStop the project. Provider jobs already accepted are not cancelled. Calls still queued are not run, even after you continue the project.
agent_choose
ChatGPTShow the user a choice as a picker (instead of asking in plain text): advertising types, templates, storyboard options, presenters, final takes (present takes with request_review first: only those can be delivered). options: [{label, asset_id?}]: an option with asset_id (for a presenter, its preview_asset_id from list_human_models) shows as an image card, one without as a button (e.g. "you choose"). The user's pick comes back as their chat message. A pick answers a question; it is not the storyboard approval, which the user gives in their own reply.
agent_finish
ChatGPTDeliver the output the user chose. asset_ids: the chosen one first (it must have been presented with request_review; when user_words names one, as "(asset_id=<id>)" or on older cards "(결과 <id>)", that one). user_words: the user's message choosing it as written, typed or sent with the card's button. idempotency_key: optional; pass a new one to try again after a refusal you have fixed (e.g. after request_review).
agent_list_runs
ChatGPTThe user's meeaww projects, newest first (id, title, status). Use when they want to continue one.
agent_resume
ChatGPTSend again the call a credit stop or a provider failure held (after the user topped up).
agent_show
ChatGPTShow the user images or videos of this project (the storyboard, finished takes, product photos) in a viewer. Tool results you receive are not visible to the user: call this for anything they should see before deciding. asset_ids: from the tool result (storyboard_ids, asset.id, image_assets).
agent_skip
ChatGPTDrop a call that is blocked (credits, an error) so the project goes on without it. Refused, and the call kept, when it may already have charged or handed a job to a provider.
agent_start_run
ChatGPTStart a meeaww production project (a run) that you direct: an ad (purpose 'ad') or a general video ('video'). The default first step for any ad or video request. text = the user's request in their words. Returns run_id for every other tool and director_brief, the rules for directing it: follow them. Free. Paid steps run against decisions the user made in chat; photos the user has go in with agent_add_images. idempotency_key: reuse it if you retry, so a retry does not open a second project.
agent_status
ChatGPTThe project's state: status, storyboard plan (approval_id, credits), assets with links, recent calls (call_id, tool, status — pass the call_id to agent_wait or agent_skip), billing block. Use it to pick up a project again; in a new chat pass include_brief=true to get director_brief, the rules for directing it.
agent_wait
ChatGPTWait for a call that returned a call_id (video takes take minutes). Returns its result when done, or its state; call again while it is queued, running or waiting_provider. Never resend the call. A finished take you show the user with agent_show.
assemble_edit
ChatGPTTrim and join existing clips in chosen order. Keep clip audio, choose one continuous source audio, or mute. Audio is reencoded, not byte-preserved. Free for the user: uses no credits. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
browse_presets
ChatGPTShow a meeaww catalog as a contact sheet with each entry's id. For retouch_image: background_preset, style_transfer, product_background_styles (then product_background with the chosen style_id). For draft_ad_image: ad_image_reference (proven ad layouts). Choose from what you actually see. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
browse_video_templates
ChatGPTList meeaww video templates: proven 15-second clips made from the user's photo(s), with a card per template (summary, best_for, needs, not_for, mood) and its default credits. You never see a template's scenario; choose by the card. Free. run_id: from agent_start_run (required: this tool reads an existing project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
compose_board
ChatGPTArrange selected images without changing their content. Layout is your choice. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
create_image_upload_url
ChatGPT로컬 이미지 파일을 i2v 입력으로 쓰기 위한 업로드 URL 을 발급한다(무과금). image_urls 에 넣을 공개 URL 이 없는 로컬 파일이 있을 때 쓴다 — t2v 로 폴백하지 말 것. 흐름: ① 이 도구로 URL 발급(이미지 여러 장이면 장당 1회) → ② 네가(에이전트가) 직접 셸에서 how_to_upload 의 curl 명령으로 PUT 업로드 → ③ image_url 을 generate_video(mode='i2v') 나 suggest_video_prompt 의 image_urls 에 넣는다. 발급된 URL 과 curl 명령은 사용자에게 보여주지 말 것(URL 소지 = 권한). jpeg/png/webp/gif/heic, 최대 20MB, 링크 유효 24시간(1회 발급으로 재사용 가능). 이미지 편집(라벨 제거 등)이 목적이고 원본이 이미 URL 로 접근 가능하면 업로드가 아니라 edit_image 를 써라 — 로컬에서 편집해 올리는 우회는 불필요하다.
crop_image
ChatGPTCrop image at pixel coordinates; original remains unchanged. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
draft_ad_image
ChatGPTPrepare a meeaww AI ad image (the app's image ad: product photos + a layout reference + your direction). First look through browse_presets kind ad_image_reference (more pages if nothing fits) and pass the layout that suits this product and message as reference_id, or pass reference_asset_id for a reference image the user gave; skip the reference only when the user asks. Free: meeaww writes the generation prompt and the on-image copy fields and returns a draft_id with the credits one image will cost. Review or edit the copy, then call generate_ad_image. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
draft_template_video
ChatGPTPrepare one template video from the user's photo(s). Free: meeaww writes the generation from the template and returns a draft_id with the credits it will cost. Ratio, resolution and language are fixed here; to change them make a new draft. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
edit_image
ChatGPTEdit supplied references; preservation is a goal, not a guarantee. Charges the user's credits like generate_image. Paid (credits). Like generate_image, planning edits need no separate consent. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
find_assets
ChatGPTList this project's actual assets and provenance. run_id: from agent_start_run (required: this tool reads an existing project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
generate_ad_image
ChatGPTCreate one ad image from a draft. Charges the user's meeaww credits (credits_per_image from the draft); failed images are refunded by meeaww. ASK mode: include the exact orders in request_storyboard_approval first and pass its approval_id. AUTO mode: proceed within the run's ad image limit. Each call is one more paid image; a resume reuses the same order. Paid (credits); runs with a recorded approval_id. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
generate_image
ChatGPTGenerate an image with optional ordered reference images. Each image made charges the user's meeaww credits by model (a failed image is not charged), storyboard frames included, so make only images the work needs. When making an advertising storyboard, develop the idea first and use the image to express its scene progression; a polished board does not validate the concept or prove product behavior. A multi-panel board must declare panels, shaped unlike the video frame. Paid (credits). Storyboard and planning images are part of making what the user asked for, with no separate consent; the storyboard approval covers takes, template videos and ad images. In an ad project a storyboard board (panels > 1) runs only after record_note. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
generate_take
ChatGPTGenerate a clip, potentially containing multiple shots. Ordered references. For a new advertisement, resolve the idea and supporting evidence before committing to production. Existing approved inputs, technical retries and narrow edits retain their execution contracts. A new call is a new paid take. Wait/resume reuses an accepted job. Paid video take (credits). Runs only with an approval_id from request_storyboard_approval that agent_approve_storyboard_plan has recorded. For an ad, get_skill ugc-production comes first. A take needs 7 to 16 minutes: if this returns call_id, keep calling agent_wait in the same turn until it is done. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
generate_template_video
ChatGPTCreate the video of a draft. Charges the user's meeaww credits (the draft's credits); meeaww refunds failed videos. ASK mode: include the draft in request_storyboard_approval(template_videos) first and pass its approval_id. AUTO mode: within the run's template limit. Each call is one more paid video; a resume reuses the same order. Paid (credits); runs with a recorded approval_id. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
get_ad_insights
ChatGPTMessaging analysis (hook, appeal points, copy structure, CTA, target signals, format) of a research's top Meta ads. Free (needs a nonzero credit balance to start new analyses); waits up to ~100s and returns what finished. Call again later for still_running or retry_later. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
get_skill
ChatGPTRead optional versioned production skills or technical references selected from optional_knowledge. User requirements take precedence. run_id: from agent_start_run (required: this tool reads an existing project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
import_product_source
ChatGPTDownloads a public HTTPS source from this product's manufacturer or official product page: the page itself (kind page; its static text is returned and saved as a document, listing the page's images), or an image/video such as an image-based detail sheet. A saved source can be cited as advertising evidence; notes and summaries cannot. Supply the source page and describe the exact product and source section. Read the returned text or image, or inspect the video, before recording claims. Source classification is agent-assessed, not proof of authenticity or efficacy. No PDFs, script-rendered content, credentials or login bypass. Other advertisers' ads and previously generated or analysis-only assets cannot be relabeled as product evidence. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
import_reference
ChatGPTDownload an accessible public HTTPS image/video URL found during research. Supply its source page and describe whether it is a thumbnail, frame or full clip. No credentials, login bypass or HTML pages. Images are shown directly; use inspect_take/extract_frame for videos. Reference appearance is inspiration, not evidence of product efficacy. Third-party ads and anything derived from them (frames, crops, boards) cannot be used as generation, edit or retouch inputs. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
import_youtube_video
ChatGPTDownload a public YouTube video (shorts included) into this run as a video asset, through the studio's fetch service. Use it when the user wants the video itself rather than a description of it: as the source_video of a take that replaces the person or product while keeping its motion and camera, as the original soundtrack of an edit (assemble_edit audio_asset_id), or for frames and trims. use=footage records that the user asked for this and vouches for the rights; use=analysis keeps the asset out of generation and edits. The file is stored at up to 1280 px with its audio and is not inspected yet. Not product evidence. At most 3 imports per run. Private, age-restricted, removed or live videos cannot be fetched. Free for the user: uses no credits. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
inspect_asset
ChatGPTRead a document or view the actual image. For video/audio use inspect_take. For an image, the result includes a written description for clients that cannot see images (container, dispenser, label text, presenter looks, each board panel): write product and presenter facts from it into prompts. Free for the user: uses no credits. run_id: from agent_start_run (required: this tool reads an existing project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
inspect_take
ChatGPTAnalyze actual video and audio (or audio alone) using Gemini. Returns timestamped observations, coverage and uncertainty, not a quality score, with the frames at the moments it reports attached as images (frames: seconds and asset ids). This run's own videos are checked only when the user asks (user_request) or for the times an edit needs (edit_points), never for defects on your own. Use prior observations for unchanged content. Only when the user asks you to check a video (purpose user_request) or an edit needs cut times (edit_points). Never on your own initiative. Free for the user: uses no credits. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
inspect_youtube
ChatGPTWatches a public YouTube video (shorts included) from its link using Gemini, without downloading it, and returns timestamped observations of shots, action, on-screen text, speech and sound for the requested interval, with uncertainty. For a YouTube link the user gives as a template or reference. It imports nothing (import_youtube_video brings the video in as an asset), and the report is not product evidence. When following someone else's video as a template, carry over structure, pacing and format, not its people, brand, exact script or music. Private, age-restricted, removed or live videos cannot be read, and a start past the video's end fails. Free for the user: uses no credits. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
list_ai_models
ChatGPTList the AI generation models this server offers: video models (Seedance and others, with each one's longest take and resolutions) and image models, the default of each flagged. Read-only and free. Call it when the user asks about models, AI models, which model is used or available, or names one such as Seedance. A real person on screen is a character (list_human_models), not a model. Answer with each model's label (for example Seedance 2.5) and what it allows. Pass the project's run_id; before a project exists, generate_take's model field lists the same video models. run_id: from agent_start_run (required: this tool reads an existing project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
list_human_models
ChatGPTBrowse active registered presenter models. Returns actual preview images and IDs; inspect before proposing/selecting. Paginated, at most six images. To let the user pick, call agent_choose with one option per candidate and its preview_asset_id as asset_id, so the user sees their photos. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
record_ad_claim
ChatGPTRecords product advertising evidence in the project: a claim, its usable wording and conditions. Each call adds a version; earlier versions stay in the project history, and later steps use the newest version of a claim_key (reuse a claim_key to correct or exclude a claim). Supported means the agent's assessment, not legal or human approval. Original before/after media must not be generated results. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
record_note
ChatGPTSave free-form direction, evidence with sources, hypotheses, or decisions. Free. In an ad project, write the creative decision here before the storyboard board and the approval request (both are refused until a note exists); write the take prompt from it. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
request_review
ChatGPTPresent actual assets for human review. Does not mark them selected. Present finished outputs: the user watches them in the viewer this opens. You have not seen a video: describe what you asked the model to make, never claim what it shows. Then let the user choose; deliver with agent_finish. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
request_storyboard_approval
ChatGPTFreezes the story, the storyboard images and the exact planned takes (and ad image orders) as one version and returns its approval_id with the credits it would cost. A changed story, reference, prompt or take makes a new version; a produced take is listed in retake_asset_ids to be made again. Nothing is generated by this call. Returns approval_id and the credit cost of the planned takes, and shows the user the board, the story and the price in a viewer. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
research_ad_market
ChatGPTCollect active paid video ads for one keyword from the Meta Ad Library (top page by reach: advertisers, recency, placements, ad copy, direct video URLs) plus, where available, the country's TikTok top ads by CTR. Free, but every call counts toward ~20 per account per day, shared with the user's own app use: do not repeat a keyword; reuse its research_id. Results are cached for 24h. Returns research_id for get_ad_insights. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
retouch_image
ChatGPTApply a meeaww photo edit, run on the user's meeaww account (their image quota). On one actual image: remove/replace/add an object, remove people, text/watermarks or clutter, restore faded color, relight, upscale; replace the background with a catalog preset (background_preset) or a description (background_from_text); place a product photo on a catalog or described background (product_background; free accounts get a watermark); apply a catalog look or another image's look (style_transfer); or a free-form instruction (instruct). text_to_image creates a new image from prompt alone. Pick catalog ids with browse_presets first; never guess them. Dedicated edits keep the rest of the photo closer than edit_image. Output is JPEG at the service's own aspect; preservation is not guaranteed. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
search_ad_reference_videos
ChatGPTSearch collected Meta video ads by keyword with paging, or list currently popular ones. Free; keyword searches are rate limited per minute. Returns direct video URLs. Keyword matching is loose: results can include ads unrelated to the product, so judge them first. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
select_human_model
ChatGPTSelect a registered presenter and import its original image. Who decides the casting follows the run purpose rules; a model the user picked is selected as is. Reuse returned asset in storyboard and video references. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
set_run_purpose
ChatGPTSet this run's purpose when it is not set: ad (advertising a product, service or brand) or video (a general, non-advertising video). The tools for that purpose appear on your next turn. It can change only until the first paid generation. Set ad or video before purpose-only tools; those return 'tool unavailable' until then. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.
transcribe
ChatGPTTranscribe original audio with word times. This does not evaluate delivery. Same purposes as inspect_take. Free for the user: uses no credits. run_id: from agent_start_run (omit to start a new project). idempotency_key: make one per intended call (required for generation, analysis and editing tools) and reuse it when retrying, so a retry never runs or charges twice. Arguments go in args.