Arckép

Cancel an in-progress chat batch.

Stop the agent's current turn. The document stays (status awaiting_input), so revise_document can follow.

Cancel a Higgsfield video that is still queued at the provider and refund the charge. Fails once the provider has started the video.

Stop the agent's current turn. The presentation stays (status awaiting_input), so revise_presentation can follow.

Run a paid chat completion. Pass model and messages (plus optional free data_guard=true to scrub PII/requisites before sending to the provider and unmask the response); streaming is unavailable in an MCP tool result.

Run free (0 RUB) Tripo v3 pre-rigging check on an owned completed 3D job (source_job_id) or public HTTPS 3D model_url.

Complete a direct multipart S3 upload for a book scan and verify its real byte size in S3.

Start a paid 3D job. Request follows POST /api/v1/models3d/generations (for model='idol', pass 1 image_url; completed job returns splat_url, ply_url, point_cloud_glb_url, preprocessed_image_url and camera).

Start a paid Direct Tripo v3 or Meshy Retexture post-processing job (rig, retarget, rig_and_retarget, texture with texture_model='tripo' | 'meshy-n1-retexture' | 'meshy-v7-retexture', segment, smartsegment, complete, decimate, uv, convert).

Create a book digitization order from a public HTTPS file_url (up to 100 MiB) or a completed multipart upload key (up to 500 MiB). Requires rights_confirmed=true (confirming copyright / permission to digitize). Nothing is charged yet: the book starts in status splitting while the worker renders pages; poll get_book until status is ready (page_count is now known), then call estimate_book and start_book (mode='sample' for a free 10-page probe or mode='full' for paid full digitization). orthography is auto (default), preserve (pre-1918 Russian: ѣ, і, ѳ) or modern. Reuse idempotency_key when retrying after a network timeout so the file is not uploaded twice.

Submit a manual correction note (kind='correction') or quality/refund review request (kind='defect') to the studio team.

Initialize a direct multipart S3 upload for a book scan up to 500 MiB (PDF or ZIP of page scans). Returns key, upload_id, part_size (16 MiB) and presigned PUT part_urls. Upload each slice via HTTP PUT, collect its ETag response header in order, and finish with complete_book_upload before calling create_book. For scans up to 100 MiB that already have a public HTTPS URL, pass file_url directly to create_book instead.

Create a paid OpenAI or Anthropic chat batch with a caller-supplied idempotency key.

Start a Word, Excel or PDF document built by Claude (POST /api/v1/documents). Paid by actual usage. The balance is charged when the agent finishes each turn; a turn that ends with no result is refunded. With too little money on the balance the answer is 402 INSUFFICIENT_BALANCE and nothing is started. This is a long task of several minutes: the answer comes at once with an id and status running. Do not wait for it in this call; fetch the result with get_document until status is awaiting_input (a turn is done) or failed. Needs format (docx = Word text document, xlsx = Excel workbook with real formulas, pdf = a finished PDF) and brief (what the document is for, who reads it, the facts and data to use: the agent invents no data). tier is fast by default (Claude Sonnet 5.5 with an Opus 5.5 advisor) or pro (Claude Opus 5.5 for the whole job, costs more): use pro only when the user asks for the best quality or the chat model is Claude Opus. Optional title and file_urls (up to 5 public HTTPS links to sources or samples: PPTX, PDF, DOCX, XLSX, TXT, MD, CSV, PNG, JPG, up to 25 MiB each and 45 MiB together; the server downloads them). A file the user gave you instead of a link is uploaded first with upload_media (a document up to 10 MiB here), and the url it returns goes into file_urls. Run estimate_document first when the user asks the price. Reuse idempotency_key on a retry after a timeout so that one document is not started twice. Tell the user the document takes a few minutes.

Start a paid Higgsfield Cinema Studio 4.0 video (POST /api/v1/higgsfield/cinema/generate). Needs prompt. duration is 4-30 seconds; resolution 480p or 720p. Reference media: image_urls (up to 30), video_urls (up to 10), audio_urls (up to 10). Each item is a public HTTPS URL or the key/url returned by upload_media (images, mp4/webm video and mp3/wav/m4a/aac audio can be uploaded). Run estimate_higgsfield_cinema first and pass its cost_rub as max_cost_rub to refuse a changed price. Optional webhook_url (public HTTPS) is called when the video completes or fails; otherwise follow the job with get_video. While it is still queued cancel_higgsfield_video cancels it and refunds the charge.

Start a paid Higgsfield Genjutsu video (POST /api/v1/higgsfield/genjutsu/{kind}/generate). kind motion/swap selects Genjutsu: source 4-30 s (swap: at least 409600 frame pixels). light-motion/light-swap selects Light: the first 15 s are used and source audio is kept. Needs video_url, image_urls (1-8), resolution 480p/720p/1080p and optional prompt. Media is a public HTTPS URL or a key/url from upload_media. Light charges input plus rounded output seconds at List price. Run estimate_higgsfield_genjutsu first; pass cost_rub as max_cost_rub. Optional webhook_url (public HTTPS) is called when the video completes or fails; otherwise follow the job with get_video. While it is still queued cancel_higgsfield_video cancels it and refunds the charge.

Start a paid Higgsfield Genjutsu Restyle video (POST /api/v1/higgsfield/genjutsu/restyle/generate). Redraws the whole source video in one chosen style; motion, camera and the source audio stay. Needs video_url (a source video of 4-30 seconds) and preset_id (an id from list_higgsfield_restyle_styles). Optional: image_urls (up to 5 character photos), prompt, resolution 480p, 720p or 1080p. Media is a public HTTPS URL or the key/url returned by upload_media. Run estimate_higgsfield_restyle first and pass its cost_rub as max_cost_rub. Optional webhook_url (public HTTPS) is called when the video completes or fails; otherwise follow the job with get_video. While it is still queued cancel_higgsfield_video cancels it and refunds the charge.

Start a paid Wan 3.0 Prime video (POST /api/v1/higgsfield/wan-prime/generate). Needs prompt. duration is 2-30 seconds; resolution 480p, 720p (default) or 1080p; aspect_ratio 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive; generate_audio (default true); enable_thinking makes the model plan the scene first (same price); optional seed. The input picks the mode: no files is text to video; image_url (the first frame, optionally with end_image_url) is image to video; image_urls (up to 10), video_urls (up to 5, 30 s together, each at least 1 s) and audio_urls (up to 5, WAV or MP3, 30 s together) are reference to video. Frames and references cannot be combined. Each file is a public HTTPS URL or the key/url returned by upload_media. Run estimate_higgsfield_wan_prime first and pass its cost_rub as max_cost_rub to refuse a changed price. Optional webhook_url (public HTTPS) is called when the video completes or fails; otherwise follow the job with get_video. While it is still queued cancel_higgsfield_video cancels it and refunds the charge.

Start a paid image job. Request needs model and prompt; max_cost_rub is optional.

Start paid LightX recamera on an owned source video. First call estimate_lightx_recamera with the same source video and camera settings. max_cost_rub is required and must cover the maximum hold. Original source audio is preserved. Returns a job ID; read get_video. Even a completed job/webhook may have is_hold=true: poll get_video until is_hold=false for final cost.

Start paid LightX relight on an owned source video. First call estimate_lightx_relight with the same source video and lighting settings. max_cost_rub is required and must cover the maximum hold. Original source audio is preserved. Returns a job ID; read get_video. Even a completed job/webhook may have is_hold=true: poll get_video until is_hold=false for final cost.

Start a pack of product cards for Wildberries or Ozon (POST /api/v1/marketplace-cards/generations). Paid. From 1-8 photos of one product and its description the service plans the pack and draws a cleaned product photo plus 3-12 slides (cover, benefits, specs, details, ...) in the marketplace's 3:4 format, keeping the marketplace's safe zones free. This is a long task of 3-8 minutes: the answer comes at once with an id and status queued. Do not wait for it in this call; fetch the result with get_marketplace_cards until status is completed or failed. Needs image_urls (public HTTPS links to JPEG, PNG or WebP photos, up to 10 MiB each; a photo the user gave you instead of a link is uploaded first with upload_media, and the url it returns goes here) and description (what the product is and the facts the cards may use: the service invents no data). Optional title and marketplace (wb is the default, ozon for Ozon). Run estimate_marketplace_cards first when the user asks the price. The pack is charged once the plan is ready, a failed pack is refunded; with too little money on the balance the answer is 402 INSUFFICIENT_BALANCE and nothing is started. Reuse idempotency_key on a retry after a timeout so that one pack is not started twice. Tell the user the pack takes a few minutes.

Start a PowerPoint presentation built by Claude (POST /api/v1/presentations). Paid by actual usage. The balance is charged when the agent finishes each turn; a turn that ends with no result is refunded. With too little money on the balance the answer is 402 INSUFFICIENT_BALANCE and nothing is started. This is a long task of several minutes: the answer comes at once with an id and status running. Do not wait for it in this call; fetch the result with get_presentation until status is awaiting_input (a turn is done) or failed. Needs brief (the topic, audience, goal and facts to use). tier is fast by default (Claude Sonnet 5.5 with an Opus 5.5 advisor) or pro (Claude Opus 5.5 for the whole deck, costs more): use pro only when the user asks for the best quality or the chat model is Claude Opus. Optional title and file_urls (up to 5 public HTTPS links to samples or sources: PPTX, PDF, DOCX, XLSX, TXT, MD, CSV, PNG, JPG, up to 25 MiB each and 45 MiB together; the server downloads them). A file the user gave you instead of a link is uploaded first with upload_media (a document up to 10 MiB here), and the url it returns goes into file_urls. Run estimate_presentation first when the user asks the price. If the result has promo_free true the deck was free for this user (say so; list_cost_rub is what it would have cost). Reuse idempotency_key on a retry after a timeout so that one deck is not started twice. Tell the user the deck takes a few minutes.

Generate Gemini speech synchronously, returning a completed result with a WAV URL. Supports style, two-speaker turns and owned custom voices. Reuse idempotency_key on retry.

Transcribe base64 audio, up to 10 MiB. Larger files use the HTTP multipart API. Completes synchronously. Keep the same idempotency_key on retries after a timeout. request contains model, language, diarize, word_timestamps and response_format.

Start a paid video job. Request follows POST /api/v1/videos/generations. For minimax-relight use create_video_relight; for minimax-camera-controls use create_video_camera; for lightx-recamera / lightx-relight use create_lightx_recamera / create_lightx_relight. Not for higgsfield-* models: use create_higgsfield_cinema / create_higgsfield_genjutsu / create_higgsfield_restyle / create_higgsfield_wan_prime.

Start paid MiniMax camera-controls on an owned starting image. First call estimate_video_camera with the same image and camera settings. max_cost_rub is required and must cover the maximum (duration + 1)-second hold. Aspect ratio follows the starting image and native audio is generated. Returns a job ID; read get_video. Even a completed job/webhook may have is_hold=true: poll get_video until is_hold=false for final cost.

Start paid MiniMax relight using an owned video and an owned lighting-sphere reference image. First call estimate_video_relight with the same media and settings. max_cost_rub is required and must cover the maximum hold. No camera, LightX, duration, audio or safety controls are accepted. Original audio is preserved; prompt is stored in history only. Returns a job ID; read get_video. Even a completed job/webhook may have is_hold=true: poll get_video until is_hold=false for final cost.

Ask a typed decision: Jev (text state + questions object) or GPT-6 Luna (input text/base64 images + questions list). Returns probabilities, not prose. Charges input tokens.

Delete one of your own custom voices (id from list_my_speech_voices). A stored voice is also removed at Google.

Create a custom voice from a text description. Free; the voice lives 365 days. request: display_name, prompt (10-1000 characters: age, timbre, pace, manner, accent), gender (male, female or neutral) and optional language_code (default ru-RU). The result holds the voice (use its id as custom_voice_id in create_speech) and a short sample. An account keeps at most 5 saved voices (made here or in the studio; expired ones do not count): the sixth fails with 409 VOICE_LIMIT_REACHED, details.limit and details.used, and delete_speech_voice makes room. Google also caps the whole project at 200 voices (the same 409). One creation per account at a time: a parallel one fails with 429 VOICE_CREATION_IN_PROGRESS, wait and retry. There is no idempotency key: after a timeout check list_my_speech_voices before retrying, and remove an extra voice with delete_speech_voice.

Download a completed 3D job as base64 binary STL or 3MF, without new provider billing.

Quote a 3D job. Pass the API model, prompt or image_url(s) and quality settings (rodin-v2.5: variant, high_pack, tpose; idol: image_url, image_type, num_gaussians, num_inference_steps, guidance_scale, seed, enable_safety_checker).

Quote exact RUB price for Direct Tripo v3 or Meshy Retexture (texture_model='tripo' | 'meshy-n1-retexture' | 'meshy-v7-retexture') rigging, animation retargeting, re-texturing, segmentation, retopology or conversion.

Quote the current per-page rate and full book price in RUB, with account discount, without charging. Pass book_id once get_book reaches status ready/sampled to quote the exact page_count and see the sample verdict (take, complex or reject), or pass page_count before uploading a scan.

Typical cost of a Word, Excel or PDF document in RUB; nothing is charged and nothing is started. Documents are priced by actual usage (the real Claude session cost times rate_rub_per_usd), so there is no exact quote: typical_cost_rub is the median final cost per tier (fast, pro) over the last 30 days, or null while there are too few finished documents. Revisions add to the cost of the same document. min_balance_rub is the balance needed to start or continue a document of that tier: below it create_document and revise_document answer 402 INSUFFICIENT_BALANCE (tell the user to top up the balance).

Quote a Higgsfield Cinema Studio 4.0 video (higgsfield-cinema-4) in RUB; nothing is charged. Same request as create_higgsfield_cinema (webhook_url is accepted and ignored). The price follows the provider's live estimate and depends on duration, resolution, audio and the reference videos you pass.

Quote a Higgsfield Genjutsu video in RUB; nothing is charged. kind motion/swap selects Genjutsu; light-motion/light-swap selects Genjutsu Light. Light uses the first 15 seconds, keeps source audio and charges input plus output seconds. Same request as create_higgsfield_genjutsu; cost_rub is the price to pass as max_cost_rub.

Quote a Higgsfield Genjutsu Restyle video in RUB; nothing is charged. Same request as create_higgsfield_restyle. The price depends on the source video length and resolution, not on the style.

Quote a Wan 3.0 Prime video (higgsfield-wan-3-prime) in RUB; nothing is charged. Same request as create_higgsfield_wan_prime (webhook_url is accepted and ignored). The price is per second of the finished video and depends on duration and resolution; sound, deep thinking and reference files do not change it.

Quote an image job. Pass API fields such as model, num_images, quality and resolution.

Quote the maximum hold for LightX recamera; nothing is charged. Required video_url is an owned source video (at least 49/24 seconds and at most 15 seconds, up to 100 MB). Pass camera='traj' with trajectory (theta, phi, radius arrays) or camera='target' with 5-element target_pose, plus recamera_mode gradual (default), bullet, direct or dolly-zoom. Original audio is preserved. Pass cost_rub as max_cost_rub when creating; bullet holds at least 10 seconds, standard modes at least 5 seconds, or ceil(source seconds), settled to actual invoice later.

Quote the maximum hold for LightX relight; nothing is charged. Required video_url is an owned source video (at least 49/24 seconds and at most 15 seconds, up to 100 MB). relit_cond_type='ic' (default) requires relight_parameters (relight_prompt, bg_source Left/Right/Top/Bottom, cfg 1-10, use_sky_mask); ref, hdr and bg require owned reference_image_url (or relit_cond_img_url) and optional ref_id (0-48). Original audio is preserved. Pass cost_rub as max_cost_rub when creating; hold covers at least 5 seconds or ceil(source seconds), settled to actual invoice later.

Price range of a Wildberries / Ozon card pack in RUB; nothing is charged and nothing is started. The planner chooses the slide count (3-12 here, 4 by default), so there is a range: cost_min_rub (1 photo, fewest slides), cost_typical_rub (1 photo, 4 slides) and cost_max_rub (8 photos, 12 slides). The pack is charged once, at its real price, when the plan is ready, and a pack that fails is refunded. A balance below cost_max_rub makes create_marketplace_cards answer 402 INSUFFICIENT_BALANCE (tell the user to top up the balance). With a personal discount the prices are the user's own and list_cost_*_rub are the price list.

Typical cost of a presentation in RUB; nothing is charged and no deck is started. Presentations are priced by actual usage (the real Claude session cost times rate_rub_per_usd), so there is no exact quote: typical_cost_rub is the median final cost per tier (fast, pro) over the last 30 days, or null while there are too few finished decks. Revisions add to the cost of the same presentation. min_balance_rub is the balance needed to start or continue a presentation of that tier: below it create_presentation and revise_presentation answer 402 INSUFFICIENT_BALANCE (tell the user to top up the balance). The answer may carry promo: a free period that applies to this caller (the chat on arckep.ru); tell the user about it in one sentence; without promo the usual price applies. image_price_rub is the price of one slide picture; pictures are charged apart from the presentation, so give it when the user asks about illustrations.

Estimate Gemini speech cost. Approximate: actual text and audio tokens determine the charge.

Estimate speech-to-text cost for a model and duration in seconds.

Quote a video job. Pass API fields such as model, duration_seconds and resolution. For minimax-relight use estimate_video_relight; for minimax-camera-controls use estimate_video_camera; for lightx-recamera / lightx-relight use estimate_lightx_recamera / estimate_lightx_relight. Not for higgsfield-* models: use estimate_higgsfield_cinema / estimate_higgsfield_genjutsu / estimate_higgsfield_restyle / estimate_higgsfield_wan_prime.

Quote the maximum hold for MiniMax camera-controls; nothing is charged. Required image_url is an owned starting image; pass prompt plus one camera control (free_camera_preset, camera_trajectory with 2-12 keyframes, camera_poses 4x4 c2w matrices, or camera_motion text). duration is 3-15 seconds (default 5); resolutions: 480p, 768p (default), 1080p. Pass the returned cost_rub as max_cost_rub when creating. The hold covers duration + 1 output seconds; final usage is settled later, refunding the difference.

Quote the maximum hold for MiniMax relight; nothing is charged. Required video_url and reference_image_url are owned uploaded media; the image is a lighting sphere. Source picture duration must be at least 56/24 seconds and at most 15 seconds (up to 100 MB). Resolutions: 480p, 768p (default), 1080p, 2k; auto aspect maps to adaptive. Original audio is preserved. Optional prompt is for history only. Pass the returned cost_rub as max_cost_rub when creating. The hold covers up to 15 output seconds; final usage is settled later, refunding the difference.

Get the status, cost and output of one 3D job.

Get the effective personal or corporate balance in RUB.

Get the status, page progress, sample verdict, RUB billing, QC summary and signed download URLs of one book. status lifecycle: splitting -> ready -> (optional sampling -> sampled) -> processing -> done (or failed). When status is done and qc.passed is true, files carries signed 24h HTTPS links for epub, fb2, docx, pdf (searchable scan + verified text layer), optional pdf_clean (typeset clean PDF), md, report_json and report_html. Only offer formats present in files. If qc.passed is false, package downloads stay closed while the team reviews the book and only report_json / report_html are exposed.

Get the structured quality report of a finished book (summary counts, detected chapters, printed TOC diff, QC checks, per-page agreement).

Get one chat batch and its provider status.

Get completed batch items and their final charges.

Get the status, cost and files of one document. status: pending/running (the agent is working: ask again in a minute or two), awaiting_input (a turn is done: result is ready and revise_document is possible), completed, failed (error says why) or archived. When result is set, give the user result.file_url (the document itself: DOCX, XLSX or PDF, see format) and result.pdf_url (a PDF copy for viewing, if there is one); result.preview_url is a picture of the first page when the agent made one; result.summary is the agent's summary (sections or sheets, assumptions, open questions) and result.changelog says what changed in this version. last_agent_message is the agent's own summary. While a revision runs, result is still the previous version. Links are signed and last about a day: call again for fresh ones.

Get the status, cost and output of one image job. Defaults to the format requested at creation. b64_json adds original file bytes to output.images (up to 16 MiB decoded in total); URLs remain present. Use response_format=url for a compact result, including when inline retrieval exceeds its limits. Reading a result never starts or charges a job.

Get the status, cost and pictures of one card pack. status: queued/processing (the pack is being built: progress_pct says how far, ask again in a minute or two), completed or failed (nothing stays charged). When completed, give the user output.images: index 0 with role clean is the cleaned product photo, the rest are the slides (cover first), each with its url, role and headline. output.warnings are machine codes such as cover_qa_failed or slide_qa_failed:2: a check the picture did not pass even after a retry; say so if the user asks why a slide looks off. Links are signed and last about a day: call again for fresh ones.

Get the status, cost and files of one presentation. status: pending/running (the agent is working: ask again in a minute or two), awaiting_input (a turn is done: result is ready and revise_presentation is possible), completed, failed (error says why) or archived. When result is set, give the user result.pptx_url (editable PowerPoint) and result.pdf_url, and show result.slides (n, title, image_url); result.deck is the presentation map (goal, audience, theme, slide plan) and result.changelog says what changed in this version. last_agent_message is the agent's own summary. While a revision runs, result is still the previous version. Links are signed and last about a day: call again for fresh ones.

Get owned completed Gemini speech with a fresh WAV download URL.

Get RUB spend of the caller's API keys: today, last 7 days, the last `days` (1-90), per day and per model, plus the balance and an estimate of the days it lasts (null without spend). key_only=true counts only the key behind this connection. Read-only; charges nothing.

Get the status, cost and output of one video job, including Higgsfield, camera and relight jobs.

List the caller's 3D jobs, newest first.

List verified page texts and multi-engine agreement scores of a book (pass sample_only=true after a free sample probe).

List the caller's book digitization orders, newest first (id, title, status, page_count, pages_done, billing).

List recent public chat batches.

List the caller's documents (id, format, title, status, version, cost), newest first. Files are in get_document.

List the visual styles Higgsfield Genjutsu Restyle can apply (nothing is charged). Each item has id, name (English), name_ru and preview_url. Pass the chosen id as preset_id to estimate_higgsfield_restyle / create_higgsfield_restyle. The list is Higgsfield's current one; source says live, cache or fallback (our snapshot when Higgsfield did not answer).

List the caller's image jobs, newest first.

List the caller's finished card packs, newest first (id, marketplace, title, cost, pictures). before_id is the id of the last pack already seen. A pack that is still running is not listed: ask for it with get_marketplace_cards by the id create_marketplace_cards returned.

List image, video, 3D, chat, decision, STT and TTS models exposed by the public API. Each model includes a human-readable description and capability tags. Use modes, resolutions, aspect_ratios, duration/reference limits, output_formats and chat context_window / max_output_tokens to choose a model; null fields mean unspecified, not unlimited. Legacy image/video features may include Studio-only functionality; rely on capabilities and modes for the public API. generation_endpoint / estimate_endpoint identify the API operation. Use the corresponding estimate_* tool to validate the exact request before generating. Metadata is fetched live from the API; capabilities describe the API (MCP chat is non-streaming, media uploads/transcription are limited to 10 MiB). minimax-relight uses estimate_video_relight / create_video_relight; minimax-camera-controls uses estimate_video_camera / create_video_camera; lightx-recamera and lightx-relight use estimate_lightx_recamera / create_lightx_recamera and estimate_lightx_relight / create_lightx_relight. Models with ids starting higgsfield- are started with the dedicated Higgsfield tools, not with create_video.

List the caller's unexpired custom Google voices, created in the studio or with this API. Use a voice's id as custom_voice_id in create_speech. saved_limit is how many saved (stored) voices you may keep. Google's private voice keys are never returned.

List the caller's presentations (id, title, status, version, cost), newest first. Files are in get_presentation.

List the caller's completed Gemini speech (studio and API), newest first.

List Google's prebuilt voices without generating audio.

List the caller's video jobs, newest first.

Clone a voice from a recording. Free. Only for a voice whose owner has agreed to be cloned. source_audio_base64: the voice sample. consent_audio_base64: the same person reading the consent phrase, recorded separately. Both are WAV, 24 kHz, mono, 16-bit, up to 5 MiB each, as base64; the sample lasts 10-30 s and the consent recording 2-30 s. The consent phrase must be read word for word, in Russian: "Я являюсь владельцем этого голоса и даю согласие Google на использование этого голоса для создания модели синтетического голоса." or in English: "I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model." Google compares the words; a mismatch fails with 400 VOICE_CONSENT_REJECTED. request: display_name, consent_confirmed (must be true: the speaker agreed) and optional store (default true: the voice lives 365 days, counts toward your 5 saved voices and uses one of Google's 200 project slots; false gives a 7-day key that is neither counted nor blocked by those limits). Errors as for design_speech_voice (409 VOICE_LIMIT_REACHED, 429 VOICE_CREATION_IN_PROGRESS). There is no idempotency key: after a timeout check list_my_speech_voices before retrying, and remove an extra voice with delete_speech_voice.

Send the agent a message about a document (POST /api/v1/documents/{id}/messages). Paid by actual usage. request.text is a change request ("add a section on risks", "make the totals formulas") or a question. The agent decides from the text: it either answers and updates the summary without rebuilding the file, or rebuilds the document when asked; result.version grows only when the file changed, last_agent_message holds the answer. Optional file_urls (up to 5 public HTTPS links or urls from upload_media, same formats and limits as create_document) add new material to the running session. Answers 202 at once; follow it with get_document until awaiting_input. It is refused with 409 while the agent is still working and for a failed or archived document, and with 402 INSUFFICIENT_BALANCE when the balance is below the minimum (see estimate_document).

Send the agent a message about a presentation (POST /api/v1/presentations/{id}/messages). Paid by actual usage. request.text is a change request ("shorten slide 3", "another theme") or a question / discussion. The agent decides from the text: it either answers and updates the presentation map without rebuilding the deck, or rebuilds the deck when asked; result.version grows only when the deck changed, last_agent_message holds the answer. Optional file_urls (up to 5 public HTTPS links or urls from upload_media, same formats and limits as create_presentation) add new material to the running session. Answers 202 at once; follow it with get_presentation until awaiting_input. It is refused with 409 while the agent is still working and for a failed or archived presentation, and with 402 INSUFFICIENT_BALANCE when the balance is below the minimum (see estimate_presentation).

Scrub personal data, company names, requisites, secrets, optional commercial terms, custom_terms ([УСЛОВИЕ_N]), named custom_labels ({'ПРОЕКТ': ['Арктика-2']} -> [ПРОЕКТ_1]), custom_patterns ({'АРТИКУЛ': 'ART-\\d+'} -> [АРТИКУЛ_1]), and optional allow_terms exceptions. Free. Returns masked_text, items, summary, and data_guard_token for unmask_data or chat_completion.

Start either a free 10-page quality probe (mode='sample', up to 10 pages/day per account) or paid full digitization (mode='full'). The book must be in status ready or sampled. mode='full' charges the full book price upfront in RUB (quote the current rate and total with estimate_book; 402 INSUFFICIENT_BALANCE if balance is too low; refused if the sample verdict was reject). Reuse idempotency_key when retrying after a network timeout so the book is not started twice. Follow progress with get_book until status is sampled (inspect sample text via list_book_pages with sample_only=true) or done.

Restore original names and requisites from placeholders like [ЛИЦО_1] in the right Russian grammatical case using data_guard_token from scrub_data or chat_completion. Free.

Upload an image, video (mp4/webm), audio (mp3/wav/m4a/aac) or a document for a presentation or document job, up to 10 MiB, as base64. Returns key and url; either can be used as a media reference (for example Higgsfield image_urls, video_urls, audio_urls). The file type is checked by content, not by the header. A document (PPTX, PDF, DOCX, XLSX, TXT, MD, CSV) is for the file_urls of create_presentation, revise_presentation, create_document and revise_document: pass the returned url there within a day. The answer also has filename, the name the presentation agent sees: Cyrillic is dropped from it, so name the file in Latin letters when the name matters.

Read one web page (http/https url) and return its text as Markdown. Charges a flat RUB price per page that is read; a page that cannot be read is free.

Search the web: ranked results with title, url, snippet and a longer summary. Pass need_content=true for the page text. Charges a flat RUB price per request that returns results; an empty result is free.