Skip to main content
Glama

Kubeez

Server Details

Generate video, images, music, voiceovers, captions and dubbing with 40+ AI models.

Ownership verified
Status
Healthy
Uptime
60.0% over the last 24 hours
OAuth
Works in Glama
Last Tested
Transport
Streamable HTTP · MCP 2025-11-25
URL
ChatGPT
–
Claude
–

TDQS

A4.1/5.0

Scored across 33 tools

Disambiguation4/5

The generation tools are cleanly split by modality (media, music, sound, dialogue, dubbing, captions, separation, background removal, trim), and the read-side tools (get_generation vs get_generations vs get_status vs view_media) have distinct, well-stated roles. The only genuinely overlapping cluster is the model-choice/pricing group — plan, pick_model, get_models, and estimate all answer 'which model and what does it cost', and their boundaries are explained in prose rather than made structurally obvious.

Naming Consistency4/5

Almost everything follows a snake_case verb_noun convention (generate_media, get_upload_url, list_assets, manage_library, remove_background, play_gemini_omni_voice). A handful of bare-verb names (estimate, plan, regenerate) and one odd noun-free outlier break the pattern slightly but remain unambiguous.

Tool Count3/5

33 tools is above the comfortable range, and a few exist only to save a turn (open_gallery, open_pricing, open_studio, play_gemini_omni_voice). The breadth is genuine, though — the server spans image, video, music, TTS, dubbing, captions, separation, uploads, library management, and planning — so most tools earn their place rather than duplicating.

Completeness4/5

Coverage is strong end to end: discovery (plan, pick_model, get_models), pricing (estimate), upload, generation across media types, polling, asset/character management, and retrieval/display of results. Notable gaps are lifecycle operations on jobs themselves — no cancel/abort for a running generation and no delete/cleanup for history — which matters on a no-refund platform.

Available Tools

33 tools
estimateA
Read-onlyIdempotent
Inspect

Scope: none (any authenticated key). Preview BOTH the credit cost AND the processing time of a generation WITHOUT deducting credits.

PRE-FETCH (mandatory): resources/read kubeez://skills/billing-confirmation — Kubeez has NO refunds. The skill explains when a preview is mandatory vs optional and how to phrase the confirmation.

Pass the same model_id you'd pass to generate_media / generate_music.

Resolution-priced image models (gpt-image-2, gpt-image-2-5-flare, gpt-image-2-5-sunburst, gpt-image-2-6-flare, gpt-image-2-6-sunburst, nano-banana-2, nano-banana-2-1, nano-banana-pro, flux-3-klein, seedream-5-5): pass the same resolution ('1K' default, '2K', '4K'; seedream-5-5 takes '1.5K' default or '2K') and generation_type you will pass to generate_media; each tier has its own price. FLUX 3 Klein is a flat-priced IMAGE model (not the per-second FLUX 3 video family): pass generation_type='image-to-image' when images will be attached, since editing is priced on its own. P-Video: pass the same quality ('720p', '1080p', or a '-draft' tier). Kling 3.0 Omni: pass reference_video_seconds whenever a clip is attached; the clip picks the billed lane (References with a clip bills its dearer -video tier, no clip runs on Frames). A References clip alone bills the clip's length; with stills beside the clip it bills the duration you pick, so also pass generation_type='image-to-image' and that duration. Kling 4.0: pass reference_video_seconds (total of the reference clips) and the same elements you will pass to generate_media; a video element moves the quote onto the -video row and adds its seconds, measured by the server (it must be a Kubeez upload). Reference clips plus video elements over 30s combined are refused. Kling 4.0 Flash (kling-4-0-flash-720p) quotes the same way with a 20s combined cap; a Flash run with first_frame_url bills the no-clip row, so quote it without reference_video_seconds. Gemini Omni 1.2 Flash (gemini-omni-1-2-flash-*): pass duration for a text, image or frames run; for a video edit pass reference_video_seconds (the clip's length is the bill) or video_edit=true, which without a length quotes the 15s maximum. GPT Image 2.6 (gpt-image-2-6-economy / -flare / -sunburst): pass reference_image_count, the number of image URLs you will send in source_media_urls (duplicates count). Any count above 0 quotes the edit, which on Flare and Sunburst adds a charge per reference image; the explanation shows base + per image x count. An edit quoted without a count assumes the 16-image maximum (pessimistic). Seedream 5.5 (seedream-5-5, resolution '1.5K' or '2K'): pass reference_image_count the same way; every reference image adds its charge, the first one included, and an edit quoted without a count assumes the 10-image maximum (pessimistic). Seedream 5.0 Pro (seedream-5-pro, -edit, -2K, -edit-2K): pass reference_image_count too; the first reference image is free and every one after it adds a charge, and an edit quoted without a count assumes the 10-image maximum (pessimistic). For an animated GIF pass output_format='gif' and the same frame_count / frame_duration_ms you will pass to generate_media: GIF is text-to-image only and has its own price by frame count, not by resolution.

Per-duration video models — pass duration (seconds). Audio defaults to ON across all video models (matches generate_media's default). For Kling 3.0 this means the per-second audio surcharge (rate in get_models — don't assume) is reflected by default; pass sound=false to preview the silent rate. For Seedance 2 audio is free (same rate either way). Seedance 2 with video reference — pass reference_video_seconds (from get_upload_session.media[].duration_seconds; omitted, the quote assumes the 15s cap). P-Video with audio input — pass reference_audio_seconds. These seconds only shape the QUOTE: generate_media always bills the server's measurement of the attached file.

Returns: • credits: { estimated_credits, mode, rate_per_second?, explanation, pessimistic } • time: { estimated_seconds, source } • price_for: seconds → the COMPLETE credit cost at that length, for THIS model, when it supports more than one. Offer the user the other lengths instead of silently keeping yours — duration is the biggest lever on a per-second bill. In agent-chat, pass these straight to show_plan_card as duration_options and the card renders tappable length chips. When credits.pessimistic is true the quote assumed a ceiling and the real bill may be lower: pass the file's measured seconds (get_upload_session.media[].duration_seconds) and re-estimate.

Inputs: model (required), sound, prompt, quality, +13 more. Returns: object

generate_captionsAInspect

PRE-FETCH (mandatory): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation. Captions are billed per-second — preview cost first.

Caption a video end to end: transcribes the speech, then burns word-timed captions onto the video server-side and returns a finished MP4. Captions are styled automatically and placed to avoid faces, on-screen text, and the main subject. Scope: generate:media. USE 'auto-caption' AS THE MODEL ID EVERYWHERE — upload and estimate both take it, and estimate returns the COMPLETE price in one call. Internally the job bills two per-second lines (speech-to-text, then the burn-in render) and estimate already sums both, itemized under breakdown; do NOT add a second estimate on top or you will double-quote. The render fee is credit-checked BEFORE transcription runs, so an unaffordable job never bills a partial step. Max 30 minutes per video. CANONICAL FLOW: (1) if the user has a local file, call get_upload_url(model_id='auto-caption'). (2) after upload, get_upload_session(token) returns total_video_seconds. (3) estimate(model='auto-caption', reference_video_seconds=<total_video_seconds>) — this is the full price. (4) quote it, ask for go-ahead. (5) only then call this tool. Required: media_url — a public HTTPS URL of the uploaded video. Optional: quality — 'auto' (default, auto-detects language, cheaper) or 'best' (specify a known language for higher accuracy). When quality='best', pass language as one of en/es/fr/de/it/pt. Optional: code_switching — true if the video mixes multiple spoken languages in one track. Optional manual placement (defaults to automatic content-aware placement): caption_x / caption_y — percent offsets of the caption block's center from the video center (x negative = left, y positive = down; y=18 is the standard lower zone); caption_scale — absolute caption size multiplier. Pass these ONLY when the user positioned the captions themselves. Async: returns {generation_id, status: 'processing', credits_charged} immediately. Poll get_generation(generation_id) until the captioned video appears in outputs.

Inputs: media_url (required), quality, language, caption_x, +3 more. Returns: object

generate_dialogueAInspect

PRE-FETCH (mandatory before polling the result): resources/read kubeez://skills/polling.

Scope: generate:speech. Generate text-to-speech audio. FOUR providers, select with provider: elevenlabs (default, ElevenLabs v3), google / gemini (Google Gemini 3.1 Flash TTS), gemini-3.8 (Gemini 3.8 Flash TTS: TWO-SPEAKER DIALOGUE with a delivery style per line), or seed (Seed 1.0 Audio: prompt-driven, supports image/audio references). ROUTING (the parameter default stays elevenlabs for backward compatibility; it is NOT the recommended pick): provider='gemini-3.8' is the recommended default for narration, bulk TTS, two-person dialogue / podcasts and simple expressive reads. provider='google' when text carries free-form [square] tags, when you need explicit language_code control, or for a long single take. provider='seed' for a scene with sound effects or music, and for voice cloning. provider='elevenlabs' only when the user names an ElevenLabs voice or wants its stability controls. ElevenLabs strips tags and ships flat audio, so leaving the default on an expressive or tagged ask is the #1 mistake. Safety net: tagged text sent with the ElevenLabs default (and no explicit ElevenLabs voice) is auto-routed to Gemini, but set provider='google' yourself rather than relying on it. NOT music generation: do not use prompt, instrumental, or the Suno V6 model fields — use generate_music for songs. Required: text — the literal words and sounds to be SPOKEN, read VERBATIM. Two mistakes to avoid on EITHER provider: (1) a DESCRIPTION ('sensual vocalizations of a woman pleasuring herself' / 'soft panting') — the model just SAYS those words; (2) a bare LIST of bracketed tags ('[moaning softly], [gasping gently], [light panting]') — nothing to actually say. [bracket] performance cues are a GEMINI-ONLY feature (provider='google'): INLINE guides placed right before the words/sounds they shape, woven INTO real spoken content — e.g. text = '[about to have an orgasm] Ohhhh im going to [slight cry] cuummm, i cannot keep it anymore'. ELEVENLABS STRIPS every [tag] before synthesis (never performed), so ANY expressive / tagged / wordless-vocalization (moans, sighs, panting) ask MUST use provider='google'; on ElevenLabs text is just the plain spoken words (bare phonetic sounds like 'Mmm... ahh...' at most, no tags). For a Gemini wordless vocalization, write the sounds PHONETICALLY ('Mmm... ahh... ohhh') with inline cues between them ([moaning], [gasp], [breathing heavily]). Character/tone go in the voice choice; keep style_prompt minimal (default 'Say the following.', see below) — NOT a scene description. If you have only a vibe and no words/sounds to perform, ASK the user what should be said. voice defaults to Rachel (ElevenLabs) / Kore (Google) if omitted. Limits below are rejected with 400 and NO credit charge.

ELEVENLABS (provider='elevenlabs'): • text 5-5000 characters after audio tags are stripped. • Audio tags ([HEY], [laughs], [whispers], …) in text/previous_text/next_text are STRIPPED server-side — NOT performed. • voice: one of 26 — Rachel, Drew, Clyde, Paul, Aria, Domi, Dave, Roger, Fin, Sarah, James, Jane, Juniper, Arabella, Hope, Bradford, Reginald, Gaming, Austin, Kuon, Blondie, Priyanka, Alexandra, Monika, Mark, Grimblewood. • language_code: one of 29 ISO codes (empirically verified — upstream rejects v2's hu/no/vi/he/fa): ar, bg, cs, da, de, el, en, es, fi, fil, fr, hi, hr, id, it, ja, ko, ms, nl, pl, pt, ro, ru, sk, sv, ta, tr, uk, zh. Pass auto to default to en. • Tuning: stability (0-1, default 0.5), similarity_boost (0-1, default 0.75), style (0-1, default 0), speed (0.7-1.2, default 1.0), previous_text/next_text (context, tags stripped). Call get_limits_for_model('text-to-dialogue-v3') to confirm the current lists.

GOOGLE / GEMINI (provider='google' or 'gemini'): • text ≤4000 BYTES (UTF-8). Inline expressive tags ([sigh], [laughing], [whispering], [shouting], [extremely fast], [long pause]) are PRESERVED and performed (unlike ElevenLabs, which strips them). • voice: one of 30 — Zephyr, Puck, Charon, Kore, Fenrir, Leda, Orus, Aoede, Callirrhoe, Autonoe, Enceladus, Iapetus, Umbriel, Algenib, Despina, Erinome, Laomedeia, Achernar, Algieba, Schedar, Gacrux, Pulcherrima, Achird, Zubenelgenubi, Vindemiatrix, Sadachbia, Sadaltager, Sulafat, Alnilam, Rasalgethi. • style_prompt (Google only): a SHORT global delivery note — overall voice character / tone for a SPOKEN-narration take (e.g. 'Speak like a theatrical vampire host, slow and velvety'), ≤4000 bytes, default 'Say the following.'. Do NOT describe the SCENE or the SOUNDS here (that reads like the thing to say); per-moment performance lives in the inline [cues] woven into text. For wordless vocalizations, leave it at the default and let the inline cues + phonetic sounds carry it. text + style_prompt must be ≤8000 bytes combined. • language_code: a BCP-47 code (e.g. en-US, es-ES, ro-RO) or auto to auto-detect. ElevenLabs tuning params (stability/similarity_boost/style/speed/previous_text/next_text) are ignored. Call get_limits_for_model('text-to-dialogue-gemini') to confirm the current lists.

GEMINI 3.8 FLASH (provider='gemini-3.8'): the provider for a CONVERSATION between two voices, or a single voice whose delivery changes line by line. • Simple: text + voice (+ style_prompt = that line's delivery). • Dialogue: speakers = 1-2 items [{"voice": "Fola"}, {"voice": "Zephyr"}] (Speaker 1, Speaker 2 in order) and dialogue_turns = ordered [{"speaker": "Speaker 1", "text": "...", "style": "slow and dry"}, ...] (≤50 turns; style optional, ≤500 chars each, ≤2000 total). text is ignored when dialogue_turns is set. • filler_words=true adds natural um/uh hesitations; only with two speakers. • Tags use ANGLE brackets inline in a turn's text: , , , , , , , , , (40 official; [square] tags are NOT performed here). • voice: 70 voices grouped by use case (default Fola), e.g. Podcast: Alnilam, Brio, Jori, Laomedeia, Rami, Veda; Storyteller: Achird, Enzo, Riko, Sami, Zali, Zubenelgenubi; Commercial: Koda, Kore, Nika, Tavi, Zeno. Full list: get_limits_for_model('gemini-3-8-flash-tts'). • No language_code: the model reads the language from the text. • Limit: ≤5000 characters across all turns, INCLUDED (tags are billed like text). ElevenLabs tuning params are ignored.

SEED 1.0 AUDIO (provider='seed'): PRE-FETCH resources/read kubeez://skills/seed-audio-prompting (or get_skill('seed-audio-prompting')) BEFORE generating — it's a PROMPT-DRIVEN generative model (voices + sound effects + background music from one prose 'director's script'), and a bare line gives bare output. text is a free-form prompt (≤2048 chars), NOT a fixed voice reading verbatim; there is NO voice param and [tags]/ElevenLabs tuning do NOT apply — describe each voice (gender/age/accent/timbre/emotion) in the prose. Optional references (image and audio CANNOT be mixed): one IMAGE — image_url or base64 image_data — to guide voice/mood; OR up to 3 AUDIO references via audio_references (a list of {speaker | audio_url | audio_data}, each clip ≤30 s) for voice cloning. Tag each cloned voice IN THE PROMPT by upload order: <<TGT_SPK1>> / <<TGT_SPK2>> / <<TGT_SPK3>> is canonical (e.g. 'Marcus (..., the actor is <<TGT_SPK1>>) says: "…"'); @Voice1 / @Audio1 / @Speaker1 are also accepted and auto-normalized to <<TGT_SPKn>> server-side, so either form binds. Optional audio_config {format, sample_rate, speech_rate, loudness_rate, pitch_rate}. Billed by OUTPUT SECONDS (capped at 120 s) — read the per-second rate from get_models / estimate, don't assume.

Billing (elevenlabs/google/gemini-3.8): PER 1000 characters (decimal ≤0.3 rounds down, >0.3 rounds up; minimum 1 credit) — read the per-1000-char rate from get_models / estimate, don't assume. Requires scope generate:speech or generate:music. Returns generation_id; poll get_status(id) until status is completed and output URLs are non-null (CDN-only). Outputs appear in media_outputs with media_type=audio.

Inputs: text, speed, style, voice, +14 more. Returns: object

generate_dubbingAInspect

PRE-FETCH (mandatory): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation. This is the most expensive model on the platform and Kubeez has no refunds — quote with estimate and get the user's explicit go-ahead before you fire.

Scope: generate:media. Translates and re-voices the speech ALREADY IN a clip, and returns one dubbed .mp4 per language. It does not write music, effects or subtitles, and it does not add speech to a silent clip.

ENGINE (model, default 'sonilo-dubbing'): sonilo-dubbing 17 languages, optional ducking, 10-second billing minimum. elevenlabs-dubbing ElevenLabs, 90+ languages (Romanian included), optional source_language (default 'auto') and cloning_strength (0-10, default 7: how strongly the dubbed voices clone the speakers), a lower per-second rate, no billing minimum. Its full code list is in get_models usage_notes and the model-selection skill. Options of one engine are refused on the other. Compare both with estimate.

INPUT: video_url, https, up to 300 seconds (the provider's own ceiling, lower than the 360 the scoring models take). Local file: get_upload_url(model_id=) → upload → get_upload_session(token) → pass the returned URL. The charge follows the clip length the server measures (duration_seconds never changes it).

LANGUAGES (required, no default). sonilo-dubbing: en, zh_cn, ja, ko, pt, pt_br, es, es_419, de, fr, it, ru, th, ar, tr, vi, id. pt_br is Brazilian Portuguese and es_419 Latin American Spanish; plain pt and es are the unqualified languages, so they are a third option rather than synonyms. ar is unqualified Arabic. Romanian is NOT available on Sonilo. elevenlabs-dubbing: BCP-47 codes such as es, es-MX, pt-BR, fr-CA, ro, pl, hi, 1 to 10 per run, any case and _ or -. ASK the user which languages they want. Do not pick a set for them, because each one is charged in full.

BILLING: clip seconds MULTIPLIED BY the number of languages, at the engine's own per-second rate. Sonilo applies a 10-second minimum to that product (not per language); ElevenLabs has none. Three languages costs three times one. estimate(model=, reference_video_seconds=, variant_count=) is the quote.

DUCKING (sonilo-dubbing only): optional, free, off by default. On, the clip's own music and effects are lowered while the dubbed voice speaks; the bed is kept either way.

ASYNC: returns {generation_id, status:'pending', languages}. Poll get_status(generation_id) — the provider has NO callback, so the poll is also what settles the run and releases any language that was not delivered. Each output's metadata.language names which language it is.

Inputs: languages (required), video_url (required), model, ducking, +3 more. Returns: object

generate_mediaAInspect

Scope: generate:media. Generate images or videos using 40+ AI models, OR extend a completed Veo 3.1 video.

============================================================ STOP. BEFORE you call this tool, you MUST fetch the skills below that match the request. This is not optional. Three equivalent fetch paths — use the FIRST one your client supports: ① TOOL CALL: get_skill(name="<skill>") — bulletproof, works in every MCP client (Claude.ai web included). No fetch tool needed. ② MCP resources/read on kubeez://skills/<name> (canonical) ③ HTTPS GET on https://mcp.kubeez.com/skills/<name>.md (caveat: hosted-agent web_fetch may block model-constructed URLs — fall back to ① if it refuses) Reading is free; firing this tool blind wastes user credits.

ALWAYS: • kubeez://skills/model-selection — picking the right model. IF the user wants a post / blog / newsletter / social / ad / thumbnail: • kubeez://skills/content-creation — MANDATORY. Quality bar + propose-before-generate rule + creative-director prompting. IF the user wants a MARKETING CAMPAIGN asset (paid social ad, landing hero, brand campaign, product launch, OOH): • kubeez://skills/marketing-campaign — MANDATORY ALONGSIDE content-creation. Niche → visual genre cheat sheet so the asset matches the industry instead of looking like generic AI marketing slop. IF the user wants a CAROUSEL / multi-slide post / LinkedIn carousel / IG carousel / 'slides about X': • kubeez://skills/viral-carousel — MANDATORY ALONGSIDE content-creation. Design modes (named style or a bespoke system for the topic) + slide archetypes + prompt bar + slide-1-anchor cohesion workflow. Without it, multi-slide outputs drift into 10 unrelated AI images. IF the user wants PRODUCT photography (packshot, marketplace white, lifestyle, in-hand, flatlay, website hero banner, ad pack, try-on): • kubeez://skills/product-photoshoot — MANDATORY. Product identity lock, mode recipes, banner copy area and headroom. IF the user wants a THUMBNAIL or a video / reel / podcast COVER: • kubeez://skills/thumbnail-generation — MANDATORY. Concept frameworks, text size and safe zones. IF the user wants a CHARACTER SHEET / turnaround / a consistent character to reuse: • kubeez://skills/character-sheet — MANDATORY. Style-aware sheet template and the reuse routes. IF the user wants their LOGO on merch / packaging / signage / a brand board, or a new logo: • kubeez://skills/brand-mockups — MANDATORY. The real logo file is attached and reproduced exactly, never redrawn. IF the user has a local file or attachment (no public HTTPS URL yet): • kubeez://skills/media-upload — the 3-step upload flow. IF the model is per-duration (any video, captions, separation, ref-video, audio-driven): • kubeez://skills/billing-confirmation — preview cost before generating. Kubeez has NO refunds. IF generation_type='extend': • kubeez://skills/video-extend — Veo 3.1 extend workflow. IF the user says 'Seedance' without specifying which generation/tier (2.5 vs 2 vs fast vs mini): • kubeez://skills/seedance-2-tiers — disambiguation rules. IF model='seedance-2-5*' AND the prompt is non-trivial (2+ reference assets, a script, keyframes, a storyboard, a blockout, >15s, or dialogue): • kubeez://skills/seedance-2-5-prompting — material roles, @Image N ordering, keyframe routes, 10000-char cap. If unsure whether a skill applies → fetch it. Reading is free.

GENERATION TYPES: • text-to-image / image-to-image / text-to-video / image-to-video — the standard set. Use get_models to confirm which a given model supports. • extend — Veo 3.1 ONLY. Pass source_task_id (NOT generation_id), prompt, and extend_model. Flat-priced per tier (lite < fast < quality) — read the exact price from get_models / estimate. Skip aspect_ratio/duration/source_media_urls in this mode. See kubeez://skills/video-extend.

SEEDANCE 2.5 DRAFT MODE: quality='draft' on seedance-2-5-480p (480p only) renders a cheap preview billed like a normal 480p render. To promote a COMPLETED draft to its 1080p final, call again with draft_generation_id=<the draft's generation_id> (prompt/model are ignored; the final reuses the draft's prompt, references, duration and ratio). Finals bill on the seedance-2-5-1080p row, read the rate from get_models / estimate. Drafts can be finalized for 7 days. Seedance 2.5 Flash (seedance-2-5-flash-*) has NO Draft mode: quality='draft' and draft_generation_id are refused on it.

IMAGE-TO-VIDEO vs TEXT-TO-VIDEO (READ THIS — agents pick wrong here a lot): • image-to-video means a strict frame TRANSITION between exactly TWO images: image #1 = literal first frame of the output, image #2 = literal last frame, the model interpolates the in-between. Use it ONLY when the user explicitly wants a transition / morph / start-to-end shot. • Animating a SINGLE image, or a video that just FEATURES one or more reference images (subject, brand, product), is text-to-video — pass the image(s) in source_media_urls and the model uses them as references. Seedance 2 text-to-video accepts up to 9 reference images; Kling 3.0 / Veo similar (see get_models). • Seedance 2 specifically: image-to-video requires EXACTLY 2 images. 1 or 3+ images with image-to-video are auto-routed to text-to-video as references. If the user gave you N images and didn't say 'transition' / 'morph from A to B' / 'start frame to end frame' — use text-to-video. • Seedance 2.5 DIFFERS from the 2.0 rule above: image-to-video takes 1 image (literal first frame) or 2 (first + last) — it is NOT auto-routed. 3+ images with image-to-video is a hard too_many_images rejection, not a silent downroute. In that mode aspect_ratio is forced to adaptive (output inherits the first frame's ratio) and NO other reference images/videos are accepted alongside the frames (reference audio still is). Text-to-video on 2.5 accepts up to 30 reference images / 10 videos / 10 audio. Seedance 2.5 Flash (seedance-2-5-flash-480p / -720p, 480p and 720p only) follows the same 2.5 rules with up to 10 reference images / 5 videos / 5 audio. To lock a first frame AND keep extra references, use text-to-video and declare the frame in the prompt — see kubeez://skills/seedance-2-5-prompting.

MODEL IDS: use only ids that came from get_models or plan — never invent or guess one. You do NOT need to call get_models every time: if you already know the id and its limits, go straight to generating. Call it when you are unsure which model fits, when you need a concrete variant, or when a capability matters (prompt_max_chars, duration_options, max_input_images, supports_negative_prompt, supports_sound, video_audio) — each row carries those plus usage_notes, and they must be respected.

VARIANT DISAMBIGUATION: Some families (seedance-2, seedance-2-fast, seedance-2-mini, seedance-2-5, seedance-2-5-flash, p-video, flux-3, gemini-omni-video, gemini-omni-flash, gemini-omni-1-2-flash, kling-3-0-omni, kling-4-0, kling-4-0-flash, kling-3-0-turbo, kling-3-0-motion-control, kling-2-6-motion-control, wan-2-7, wan-3-0, grok-imagine-video-1-5-preview) expose multiple concrete model_ids (resolution tiers, draft vs standard, with/without video-ref). PREFER passing a concrete model_id up front: when the user gave no resolution preference, default straight to the recommended variant (e.g. seedance-2-720p) instead of the bare family, so there is NO round-trip. If you DO call with a bare family label, the server returns variant_required with a recommended_variant: ADOPT it and continue straight to estimate / show_plan_card, where the cost (and thus the resolution/price trade-off) is confirmed, so a separate variant question is usually unnecessary. Ask the user only when they signalled they care about price/quality, or when the applicable variants differ in CAPABILITY (not just resolution). EXCEPTION: when a reference VIDEO is attached, pass the BARE family (do NOT hand-pick a concrete id) and the server selects the correct -video-ref variant and its billing for you.

ASPECT RATIO: ALWAYS use the aspect_ratio parameter — do NOT bury 9:16 / 16:9 / 1:1 in the prompt. 'banner'/'landscape' → 16:9, 'vertical'/'story' → 9:16, 'square' → 1:1. The prompt should describe content, not dimensions.

GPT-IMAGE-2 — 2K and 4K REQUIRE a non-square, non-auto aspect_ratio (9:16 / 16:9 / 4:3 / 3:4). Calling 2K/4K with auto or 1:1 returns 400 and credits are NOT held.

GPT-IMAGE-2-6: three model_ids: gpt-image-2-6-economy (1K only, 8 aspect ratios, no 3:4 / 4:5), gpt-image-2-6-flare and gpt-image-2-6-sunburst (1K / 2K / 4K, 16 aspect ratios). Up to 16 reference images. An edit (any source_media_urls) costs the edit row price plus a charge per reference image on Flare and Sunburst (duplicates count), so estimate it with reference_image_count. Optional background='transparent'|'opaque'|'auto' at 1K only. Optional output_format: 'png' (default), 'jpeg' / 'webp' (Flare and Sunburst), or 'gif' for an animated GIF. GIF is text-to-image only (refused with source_media_urls) and has its own price: Flare and Sunburst need frame_count, a whole number 2 to 16; Economy takes frame_count 'auto', 4, 9, 16, 25, 36 or 64 (optional). Optional frame_duration_ms 20 to 2000 per frame. Read GIF prices with estimate(output_format='gif', frame_count=N). An unsupported resolution, aspect_ratio, output_format or frame_count is refused before credits move.

SEEDREAM-5-5: two rows, seedream-5-5 (1.5K) and seedream-5-5-2K, or pass resolution='1.5K' (default) / '2K' on seedream-5-5. Text-to-image with no source_media_urls, or edit with 1 to 10 reference images on the same id (no -edit ids). An edit costs the row price plus a charge per reference image, the first one included (duplicates count), so estimate it with reference_image_count. aspect_ratio 1:1, 4:3, 3:4, 16:9 or 9:16 (no auto). Anything else is refused before credits move.

AUDIO ON VIDEO: the sound parameter defaults to TRUE, so audio is ON for EVERY model that supports it unless the user explicitly asks for a silent clip (only then pass sound=false). Read capabilities.video_audio: 'included' means audio is always bundled and the sound param is ignored; 'toggle_via_sound_param' means the sound flag controls audio and DEFAULTS ON (free on Seedance 2; on Kling 2.6, Kling 3.0 and Kling 3.0 Omni Frames / References (their -audio rows cost more per second at 720p and 1080p) audio costs extra, so keep it on by default but state the higher audio price in your cost quote so the user is not surprised); 'silent' means no generated audio (Seedance 1.0 only). Never silence a model's audio on your own initiative. Do NOT infer audio from supports_sound alone.

CONSISTENT CHARACTER + VOICE (Google Gemini Omni Video): this is the ONE family that can lock a REUSABLE on-camera identity. Every other talking-video model (Veo 3.1, Kling 3.0 Pro, P-Video) re-rolls a fresh face and voice on each generation, so a multi-clip campaign drifts. Gemini Omni Video does NOT: save a character once, then every clip renders the SAME person with the SAME voice. WHEN TO REACH FOR IT (and proactively suggest it): the user wants a recurring spokesperson / mascot / brand presenter, a video SERIES or ad set, or anything where 'keep the same person/voice across clips' matters. WORKFLOW: (1) manage_library(kind='character', action='create', descriptions=..., image_url=..., character_name=..., voice_id=...) saves the identity once (bind a voice_id from list_gemini_omni_voices so the voice is locked too); (2) for each clip call generate_media with a gemini-omni-video-* / gemini-omni-flash-* model, passing character_ids=[<character_id from list_gemini_omni_characters>] and/or voice_id=.... character_ids and voice_id are accepted ONLY by gemini-omni-video-* / gemini-omni-flash-* models (ignored elsewhere; Gemini Omni 1.2 Flash, gemini-omni-1-2-flash-*, has no characters or voices and refuses them). Each character_id consumes 1 of 7 reference slots. This is VIDEO identity; it does NOT change the image-model rule (character consistency in still images stays on gpt-image-2-5-sunburst).

KLING 4.0 ELEMENTS + VOICES (kling-4-0-* ONLY, refused on every other model): elements = up to 7 one-time subjects [{name, image_urls (2-4) OR video_url (one 3-8s MP4/MOV character clip uploaded to Kubeez), tag?, description?, voice_url?}], voices = up to 3 [{name, audio_url}]. Voice samples are 5-30s; reference clips (3-30s each) plus video elements are 30s combined. Mention each in the prompt as @name; unmentioned ones are not sent, an unmatched @ is refused. A video element bills like a reference clip (the -video row plus its seconds), so pass the same elements to estimate. Full rules: get_models usage_notes for kling-4-0. list_assets returns saved Kling 4.0 elements and voices already in this shape; save new ones for reuse (they also appear in the user's studio) with manage_library(kind='kling_element' | 'kling_voice', action='create'). KLING 4.0 FLASH (kling-4-0-flash-720p): a separate Kling 4.0 tier, 720p only, 3-20s. Same references and elements (reference clips 2-20s each, clips plus video elements 20s combined) but NO voices (bind a voice_url on a character element instead). Saved Kling 4.0 elements work on Flash too; saved voices do not. KLING 4.0 FRAMES (start + end frame): first_frame_url (@first_frame) starts the video; on kling-4-0-720p / -1080p add up to 8 keyframe_urls in order (@keyframe_1 ...) and a closing last_frame_url (@last_frame), frames in total at most min(10, seconds + 1) counting first and last. Flash takes the first frame only. Frames exclude references and elements, follow the first frame's shape (no aspect_ratio) and bill the clip-free row.

KLING 3.0 STORYBOARD + ELEMENTS (kling-3-0-std / -pro / -4k ONLY): shots = 1-5 [{prompt (<= 500 chars, 2500 in total), duration (whole seconds 1-12)}], 3-15s in total; the video is as long as the shots added up and bills those seconds, needs sound on, and prompt is required but not sent (pass a one-line summary). elements on Kling 3.0 = up to 3 [{name (letters, digits, _), description, image_urls (2-4 images)}], mentioned as @name in the prompt or a shot; every @word must be a passed element, and a mention needs a start frame in source_media_urls. Image elements add no cost.

STUDIO SWITCHES (each refused on any other model, unset = the model's default): video_edit=true (Seedance 2.5 and Seedance 2.5 Flash: edit the one attached reference clip; the output keeps its length and frame, no duration; billed on the -video-ref row with the clip's measured length as the output seconds, estimate with video_edit=true and reference_video_seconds; also accepted on gemini-omni-1-2-flash-*, where a clip alone already makes an edit), fps 24 | 48 (P-Video, default 48), enable_translation (Veo 3.1, default on), prompt_extend (Wan 2.7, default on), turbo (p-image-edit, default on), output_format png | jpeg (nano-banana, nano-banana-edit, nano-banana-2, nano-banana-2-1, nano-banana-pro, qwen-text-to-image, qwen-image-to-image; default png).

GEMINI OMNI 1.2 FLASH (gemini-omni-1-2-flash-360p / -720p / -1080p / -4k, the id sets the resolution): per-second, built-in audio, a separate model from 1.1 Flash. The mode follows what you attach: prompt only or up to 6 images in source_media_urls (duration 3-15); first_frame_url alone (3-15); first_frame_url + last_frame_url (3-10; frames exclude images and video); or ONE clip in source_media_urls / reference_video_urls (at most 15s, a Kubeez upload, up to 5 images beside it) = a video edit with NO duration, billed on the clip's measured length. No trim window (trim_video first), no character_ids, voice_id or audio (use 1.1 Flash for those). aspect_ratio auto (default), 16:9 or 9:16; seed 10000-99999. Full rules: get_models usage_notes.

Inputs: model (required), prompt (required), fps, mode, +38 more. Returns: object

generate_musicAInspect

PRE-FETCH (mandatory before polling the result): resources/read kubeez://skills/polling — first-poll delays + cadence.

Scope: generate:music. Not generate:media — use generate_media for image/video, this tool only for music. Generate AI music. DEFAULT to SIMPLE mode for all prompt-based requests. Only switch to ADVANCED when the user explicitly provides or asks for lyrics, a specific song title, style tags, or vocal gender. SIMPLE MODE (default): set prompt only (max 400 chars); leave title, style, vocal_gender, song_description, and lyrics empty. Optional: instrumental, model. ADVANCED MODE (custom, use only when user asks for lyrics/title/style/vocal control): set BOTH title AND style (non-empty). Required: vocal_gender = m, f, or random (random = provider picks). Content: exactly ONE of song_description OR lyrics (never both). ADVANCED VOCAL songs need explicit lyrics — in custom mode the provider sings the text VERBATIM, so a song_description there gets auto-routed back to SIMPLE mode (style/title folded into the prompt; response carries mode_note). If you only have a description, use SIMPLE mode directly. Advanced instrumental: song_description only (omit lyrics). Optional: negative_tags, style_weight, weirdness_constraint, audio_weight (default 0.65). Invalid: title without style or style without title. MODEL: OMIT model so the server auto-picks the newest, best engine (currently V6). Only set it when the user explicitly names an engine. Valid: V6 (richest detail, most natural vocals), V6_WILD (bolder, more distinctive), V6_MINI (fast and light). Every version up to V5_5 is discontinued by the provider and is silently remapped to V6 at the same price. TREBLO: model='treblo-v3' is a second engine (full songs, vocal or instrumental, writes its own lyrics and picks its own length). It takes prompt (or song_description), optional lyrics, style, title, negative_tags and instrumental, with no simple/advanced split and no title+style pairing rule. vocal_gender m/f is optional and becomes its 'male vocalist'/'female vocalist' style tag; the three weights do not apply (ignored, reported in mode_note); persona_id and exact_duration_ms are refused. Style is split on commas into tags and the engine only takes tags from its own vocabulary: use short genre/era/vocal words ('pop, rap, 2010s'), not tempo, key or sentences ('100 BPM', 'A minor', 'catchy party hook'). Words it does not take are moved into the description or dropped, and named in mode_note; put tempo, key and mood phrasing in the prompt. REAL ARTIST NAMES are accepted as a style reference, in the prompt ('in the exact style of The Weeknd, Starboy era, ...') and as a style tag ('the weeknd, synthwave, pop, male vocalist'); the vocal is generated, not the artist's own voice. When a user asks for a song in the style of a named artist, use treblo-v3 and pass the name through: do not strip it or swap it for a genre description. Add 'No artist names in the lyrics.' to the prompt, or the lyrics can sing the name. Never paste a real song's lyrics. Poll get_status the same way. PERSONA: to reuse a saved persona, pass persona_id (from generate_music_persona or list_music_personas). It copies the singer's VOICE by default (persona_model='voice_persona', supported across the V6 family). Pass persona_model='style_persona' to copy the genre/vibe instead. Applying a persona forces custom mode automatically. LENGTH: Suno picks its own. For an instrumental bed at a length you choose, call generate_sound(model='sonilo-text-to-music', duration_seconds=...) instead — a length passed here is served there anyway, and cannot be combined with lyrics/vocal_gender/persona_id. Returns generation_id; poll get_status(id) until completed.

Inputs: model, style, title, lyrics, +12 more. Returns: object

generate_music_personaA
Idempotent
Inspect

Scope: generate:music. Create a reusable PERSONA from a song you already generated, capturing its voice and style. Synchronous and FREE: returns the persona in one call (no polling). Inputs: song_id (a music_songs id from get_status on a completed music generation), name (<=80 chars), description (<=500 chars; the voice/mood/genre to capture), optional style tag. Requirements: the source song must be fully generated on a model above v3.5 (all current models qualify); each song can only have ONE persona (a repeat call returns the existing one). Use the returned persona_id with generate_music(persona_id=...) to sing new songs in that identity.

Inputs: name (required), song_id (required), description (required), style. Returns: object

generate_separationAInspect

PRE-FETCH (mandatory): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation. Separation is billed per-minute of audio — preview cost first.

Split an audio track into vocals + instrumental stems (MVSEP BS Roformer). Scope: generate:media. Billing: per-minute of audio duration, measured by the server; the audio must be a Kubeez upload. CANONICAL FLOW: (1) if the user has a local file, call get_upload_url(model_id='mvsep-40') and send the link. (2) when they confirm upload, call get_upload_session(token) — it returns total_audio_seconds. (3) call estimate(model='mvsep-40', reference_audio_seconds=<total_audio_seconds>) to get the exact price. (4) tell the user the price and ask for confirmation. (5) ONLY THEN call this tool. Don't charge silently. Required: media_url — a public HTTPS URL of the uploaded audio. Async: returns {separation_id, task_id, status: 'pending'} immediately. Poll get_status(separation_id) until status is 'completed'. When completed, the response carries vocals_url + instrumental_url (permanent CDN). For a full workflow with preview + library, direct the user to kubeez.com/audio/separation.

Inputs: media_url (required). Returns: object

generate_soundAInspect

PRE-FETCH (mandatory for the video models): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation. Every model here bills PER SECOND and Kubeez has no refunds — quote with estimate before you fire.

Scope: generate:media. Sound effects and music, either from a sentence or written against a clip you already have. NOT for songs with vocals — that is generate_music (Suno).

PICK BY WHAT YOU HAVE: You have a sentence: sonilo-text-to-sfx one described sound effect. 1-180s, default 5s. 3s billing floor. sonilo-text-to-music an instrumental track. 5-360s, default 60s. 10s billing floor. You have a clip (video_url REQUIRED, https, up to 360s): sonilo-video-to-sfx effects matched to the picture, as a bare audio track sonilo-video-to-music a score written to the cut, as a bare audio track sonilo-video-to-sound both, mixed, as a bare audio track sonilo-video-to-video-sfx same as above but returns the CLIP with the sound in it sonilo-video-to-video-music same, muxed sonilo-video-to-video-sound same, muxed Prefer a -video-to-video-* id when the user wants a finished clip back; prefer the bare audio ids when they will mix it themselves in an editor.

PROMPTS: pass prompt and it is folded into the field that model reads. The two *-sound models write effects AND music, so they take music_prompt and sfx_prompt separately and reject a lone prompt. Max 2000 chars each. A video run with no prompt at all is valid — the model writes to what it sees.

BILLING: per second of the CLIP, not of the output, and each product carries the provider's OWN minimum billable length: 10s on anything that writes MUSIC (both mixed *-sound ids included), 3s on the effects-only ids. A 6s score bills 10s; a 2s effect bills 3s. estimate(model=..., reference_video_seconds=) applies it.

CLIP LENGTH: a video run bills the clip's length as the SERVER measures it, so the clip must be a Kubeez upload (get_upload_url) or one of your own Kubeez generations; any other URL is refused. duration_seconds does not change the bill on a video run.

LOCAL FILE: get_upload_url(model_id='sonilo-video-to-video-sfx') → upload → get_upload_session(token) → pass the returned URL as video_url. Its media[].duration_seconds is the measured length, for an exact estimate.

ASYNC: returns {generation_id, status:'pending'}. Poll get_status(generation_id). Sonilo has NO provider callback, so the poll is also what settles the run — keep polling until the status is completed or failed, or the credits stay reserved.

Inputs: model (required), prompt, ducking, video_url, +4 more. Returns: object

get_balanceA
Read-onlyIdempotent
Inspect

Scope: read:balance. Check the user's current credit balance. Always call this before generate_media or generate_music to avoid 'insufficient credits' errors. Returns available credits; each generation consumes credits based on the model (see get_models for cost_per_generation).

Inputs: no named parameters. Returns: object

get_generationA
Read-onlyIdempotent
Inspect

Scope: read:generations. Fetch a SINGLE past generation by id — full detail (prompt, model, params, status, outputs, credits_deducted, error_message). Use when you have a specific id and want one row; vs get_generations which returns a paginated history. Pairs with regenerate(id, ...) for iteration: get_generation(id) → inspect params → regenerate(id, prompt='tweaked'). Returns the same row shape as one entry of get_generations.generations[].

Inputs: id (required). Returns: object

get_generationsA
Read-onlyIdempotent
Inspect

Scope: read:generations. Return the user's media-generation history, newest first — covers BOTH the current session's jobs (still queued / processing / just-completed) AND prior sessions. Each row includes id, prompt, model, generation_type, status, created_at, completed_at, processing_time_ms, credits_deducted, error_message, and an outputs[] array with the playable/viewable URL(s), thumbnails, dimensions, duration, format, and file size — enough to render a gallery without a follow-up call. limit is 1–25 (default 25; higher values are clamped). Optional filters: status, model (model_id), generation_type (text-to-image, image-to-video, etc.). Use this when the user asks 'what have I made', 'show me my last 10 videos', or to surface re-usable assets. For a single generation's full live status (poll until done), use get_status(id).

Inputs: limit, model, status, generation_type. Returns: object

get_modelsA
Read-onlyIdempotent
Inspect

Scope: none (any authenticated key). List available AI models with costs, limits, and input requirements. Call it when you are unsure a model or capability exists, or when comparing models — use only model_id values returned here. NARROW THE LIST, do not read all of it: the unfiltered video catalog is over 140 rows. accepts='image'|'video'|'audio'|'document' keeps only models that take that input type. duration_s= keeps only models that can render that clip length. sort='price' orders cheapest-first at that duration (default 'id'). verbose=false drops usage_notes / cost_note / estimated_time_seconds for a compact pick list. Example, animating a logo into an 8-second clip: get_models(model_type='video', accepts='image', duration_s=8, sort='price', verbose=false). PRICE: compare models on price_for, never on cost_per_generation. price_for maps duration in seconds to the COMPLETE credit cost at that length (e.g. {'8': 45} = 45 credits for an 8-second render), so price_for['8'] is directly comparable across every model, and its ABSENCE means the model cannot do 8 seconds. billing_unit ('per_second' | 'per_generation' | 'per_1000_chars' | 'per_minute') and rate give the underlying unit. cost_per_generation is legacy and NOT unit-comparable — for per-second models it is a pre-multiplied example duration that differs per family. Each model also returns: model_id, duration_range_s, estimated_time_seconds (REAL — DB-authored override or historical average of past completions; use it as the first-poll delay, do NOT poll every second), requires_input_media, input_media_types (e.g. ['image'] or ['image','video']), generation_types, capabilities (prompt_max_chars, duration_options, max_input_images, max_input_videos, supports_negative_prompt, supports_sound, video_audio), and usage_notes with exact rules. For video models, video_audio is critical: 'included' = audio in output without sound param; 'toggle_via_sound_param' = sound true/false controls audio and DEFAULTS ON, so audio is on for every supporting model unless you explicitly silence it (free on Seedance 2; Kling 2.6 and Kling 3.0 charge extra for audio but still default on, so state the higher price in your quote); 'always_on_flat_rate' = audio is always produced and the price does not change with it; 'silent' = no generated audio (Seedance 1.0). Only pass sound=false when the user explicitly asks for a silent clip. Do not answer 'which models have audio' using only supports_sound — that flag only means the sound parameter exists for toggling. latest_in_family maps a family to the CHEAPEST variant of its newest release. When requires_input_media is true, you MUST pass source_media_urls with the correct count and types to generate_media. If the user has media only on their device (no URL), use get_upload_url → user uploads at link → get_upload_session before generate_media — see those tool descriptions.

Inputs: sort, accepts, verbose, duration_s, +1 more. Returns: object

get_skillA
Read-onlyIdempotent
Inspect

Scope: none (any authenticated key). Fetch the full markdown body of a kubeez skill by name. USE THIS as the universal fallback when your MCP client cannot read MCP resources (e.g. the Claude.ai web app does NOT expose resources/read to the model) or cannot fetch arbitrary HTTPS URLs (e.g. Claude.ai web's web_fetch only allows URLs that came from the user message or a prior tool/search result, so model-constructed URLs like https://mcp.kubeez.com/skills/viral-carousel.md get blocked). Tool calls always work in every MCP client, so this path is bulletproof.

AVAILABLE SKILLS: • content-creation — quality bar + propose-before-generate for posts/blogs/ads • social-post — single-image social/ad post = a DESIGNED creative (hook headline rendered on the image + caption), NOT a bare photo; 6 archetypes + prompt bar • marketing-campaign — niche-aware visual genre cheat sheet for marketing assets • viral-carousel — design modes (named style, reference, or a bespoke Art director system) + archetypes + prompt-construction bar for multi-slide IG/LinkedIn carousels (fetching via THIS tool also appends the user's recent deck history for anti-repeat — prefer it over resources/read for carousels) • carousel-style-previews — carousel style library: the 9 named styles (The Press, The Modernist, The Dossier, The Billboard, The Index, The Photograph, The Moleskine, The Whiteboard, The Screenshot) with preview URLs, plus palettes, card treatments, type pairings, mark styles and deck recipes • model-selection — picking the right image/video/audio model from get_models • media-upload — required when user has a local file (3-step upload flow) • asset-library — persistent named media (logos, references, songs) • billing-confirmation — preview cost before per-duration jobs (Kubeez has NO refunds) • polling — first-poll delays per media type, then every 5s • video-extend — extending Veo 3.1 clips via generation_type='extend' • seedance-2-tiers — Seedance 2.5 vs Seedance 2 / Fast / Mini disambiguation • seedance-2-5-prompting — writing the seedance-2-5 prompt itself: material roles, @Image N mapping, keyframes, storyboards, blockouts, long-video staging, 10000-char cap • critique — the 6-point rubric you (the calling LLM) run on your own output before shipping • marketing-strategy — hooks, platform-native copy formats, funnel/CTA matching, paid-ad angle system + character limits (the COPY twin of content-creation) • product-marketing — foundation context doc (product / audience / positioning / customer language) that copywriting, email-marketing, and launch read first • copywriting — off-image conversion copy: landing/pricing/feature pages, headlines, CTAs, value framing, 7-sweep editing pass • email-marketing — lifecycle sequences (welcome / nurture / onboarding / win-back) + B2B cold outreach and follow-ups • launch — launch playbook: announcement sizing, owned/rented/borrowed channels, five phases, checklist • anti-slop : the shared quality floor read alongside any task skill: visual and spoken slop tells, photoreal person module, negative-direction blocks, claims floor • ugc-video : creator-style / UGC / TikTok-style video: anti-polish contract, five arcs, hook gate, truth contract, frozen-frame QA • faceless-video : narrator-led video with nobody on camera: the one-call ceiling, script spine, style lock, captions pass • product-photoshoot : packshots, marketplace whites, lifestyle, in-hand, flatlay, hero banners, ad packs, try-on, conceptual product stills; product identity lock + recipes • thumbnail-generation : YouTube thumbnails and reel / podcast / video covers: concept frameworks, emotion casting, safe zones, 2 to 4 word text rules • character-sheet : turnaround / expression / model sheets as an identity anchor, plus how to reuse it (asset library, Seedance refs, Gemini Omni characters, Kling elements) • brand-mockups : the real logo on merch, packaging, signage, stationery, screens, brand boards; logo-maker when there is no logo yet • gemini-tts-expressive : expressive, tagged, accented or non-English TTS via generate_dialogue(provider='google') • seed-audio-prompting : prompt-driven speech via generate_dialogue(provider='seed'): prompt structure, voice references, audio_config

Returns: { name, uri, url, content_type, content } where content is the full skill markdown body. Same source-of-truth as the kubeez://skills/ MCP resource and the https://mcp.kubeez.com/skills/.md HTTPS endpoint — pick whichever path your runtime supports. CALL THIS BEFORE generate_media / generate_music / generate_dialogue when the matching skill applies; firing those tools without the relevant skill loaded leads to AI-slop output and wasted credits.

Inputs: name (required). Returns: object

get_statusA
Read-onlyIdempotent
Inspect

Scope: read:generations. Poll the status of ANY job — media generation, music generation, or audio separation — by id.

PRE-FETCH (mandatory on the FIRST poll of a session): resources/read kubeez://skills/polling — first-poll delays per media type, then every 5s.

Auto-dispatches: pass the id and the server figures out which kind of job it is. Optionally pass kind to skip the lookup.

Returns the merged shape: • kind ('media' | 'music' | 'separation') — what the id resolved to • status — 'pending' | 'processing' | 'completed' | 'failed' • outputs / songs / vocals_url+instrumental_url — depending on kind • error_message on failure

Output URLs are CDN/R2 only — never vendor temp links. Keep polling until status='completed' AND those URLs are non-null. To SHOW the user a finished image, call view_media(output_id) — never download the CDN URL yourself.

Inputs: id (required), kind. Returns: object

get_upload_sessionA
Read-onlyIdempotent
Inspect

Scope: generate:media. PRE-FETCH (mandatory if you haven't already): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation. The duration fields returned here are what powers exact per-second billing — handle them per the skills.

Fetch uploaded file URLs from a temp upload session created by get_upload_url. Call this after the user confirms they finished uploading at the link (or to poll until files appear). Pass the same token get_upload_url returned. Returns media_urls for use as source_media_urls in generate_media (or other tools expecting HTTPS media URLs). ALSO returns a per-file media array with {url, kind: image|video|audio, duration_seconds?, measurement?, width?, height?, mime_type?, file_size_bytes?} and aggregates total_video_seconds / total_audio_seconds / unknown_duration_count. duration_seconds is the SERVER's own measurement of the file, the exact number length-billed models charge on (e.g. seedance-2-fast-*-video-ref: (ref_video_seconds + output_seconds) × rate; p-video with audio: audio_seconds × rate). Pass it to estimate for an exact quote. billing_readiness is 'exact' only when every audio/video file is measured. If unknown_duration_count > 0, media[].measurement says why ('unavailable': call again shortly; 'unreadable': ask for a standard MP4/MOV/MP3/WAV); length-billed models refuse such a file rather than guess. If count is 0, nothing uploaded yet — remind them to use the link; you cannot substitute local files.

Inputs: token (required). Returns: object

get_upload_urlAInspect

Scope: generate:media. PRE-FETCH (mandatory): resources/read kubeez://skills/media-upload — explains the full flow, what the response contains, and how duration_seconds powers exact billing.

Create a temporary upload session that returns BOTH a browser link (for a human user) AND a direct upload endpoint the agent POSTs to itself from its own environment (shell, code execution sandbox, Claude Code, Cursor). REQUIRED whenever the user wants to supply a photo / video / audio file to edit, animate, reference, lip-sync, or stem-separate but does not already have a stable public HTTPS URL. Pass model_id from get_models for the generation you plan (enforces max_input_images and input_media_types). Response fields: • url — browser link the user opens to upload (no Kubeez login). • direct_upload_url + direct_upload_curl_example — POST endpoint for AI agents that already have the file on disk. Multipart form: token + one or MORE file parts (repeat -F "file=@..." to batch several files in one curl). Same token as the browser link; both write to the same session. • direct_upload_python_example — the same upload with Python requests, for sandboxes without curl. • direct_upload_allowlist_domains — the one domain the upload needs (a kubeez.com host). If your sandbox blocks the request (DNS failure, connection refused, proxy 403), ask the user to allow it in the sandbox's network settings, or try direct_upload_fallback_url. • token — pass to get_upload_session afterwards. If you (the agent) already have the file in your environment, upload it yourself with the direct endpoint — no need to round-trip the user. If the file is on the user's device only, share url and ask them to upload there. After upload(s), call get_upload_session(token) to retrieve media_urls, then pass those as source_media_urls to generate_media.

Inputs: model_id (required). Returns: object

list_assetsA
Read-onlyIdempotent
Inspect

Scope: generate:media. List the user's persistent Asset Library — named, reusable images / videos / audio stored in their private R2 bucket. Each entry has a stable name (e.g. tesote-logo), an id, kind (image|video|audio), size_bytes, and a freshly signed CDN url valid for 1 hour. USE WHEN: the user references a brand asset, logo, recurring face / voice / song clip, or says 'use my logo / my photo / my track' — before asking them to re-upload, check here first; their library may already have it. REFERENCING IN GENERATIONS: pass the url field of any matching asset directly into generate_media as a source_media_urls value — the URL is a public signed CDN link the model can fetch. Also returns quota_bytes / used_bytes so you can warn the user before calling manage_library(kind='asset', action='add', ...) if they're near the cap (default 50 MB). TEAM SEATS: active team members also see their team's shared assets merged into assets — those rows carry shared: true, owner_user_id, and uploaded_by {user_id, username}; a slim team block reports the team pool's quota_bytes / used_bytes. KLING ELEMENTS: when the user has saved any, a kling_elements block lists them — named reference bundles (name, description, image_urls, video_urls) they built in the app to keep a person, product or place consistent on kling-3-0-* models. Check here too before asking for a reference photo; pass any of their URLs into generate_media as source_media_urls values. Saved Kling 4.0 subjects and voices come back separately under kling_4_elements / kling_4_voices, already shaped for generate_media's elements / voices arguments on kling-4-0-* models (mention each as @name; Kling 4.0 Flash takes the elements but not the voices). Save new ones with manage_library(kind='kling_element' | 'kling_voice').

Inputs: no named parameters. Returns: object

list_gemini_omni_charactersA
Read-onlyIdempotent
Inspect

Scope: generate:media. List the user's saved Gemini Omni characters (newest first). Each character is a bundle of (reference image, description, optional voice_id) backed by an opaque provider character_id that the model uses to render the same identity across multiple clips. USE WHEN: the user references a saved character ('use my Maya 'character', 'render the brand mascot') — pull the matching character_id from this list and pass it in character_ids when calling generate_media with a gemini-omni-video-* / gemini-omni-flash-* model. ALSO check this PROACTIVELY whenever the user wants the SAME person + voice across multiple video clips (a series, a recurring spokesperson, a marketing ad set): if no fitting character exists yet, offer to create one first via manage_library(kind='character', action='create', ...) so the identity is locked across the whole campaign. RETURNS: characters (each row has id [internal UUID], character_id [the value to pass to generate_media], character_name, image_url, description, voice_id, created_at). Each character_id consumes 1 of the 7 reference slots in generate_media.

Inputs: no named parameters. Returns: object

list_gemini_omni_voicesA
Read-onlyIdempotent
Inspect

Scope: none. Return the 30-voice catalog for Gemini Omni Video. Each entry has an id (forward as voice_id to generate_media), label, gender (female / male / ungendered), and a one-line character description (e.g. 'Soft, high pitch'). USE WHEN: the user wants to pick a voice for a gemini-omni-video-* / gemini-omni-flash-* generation, or you want to suggest 2–3 fitting voices for the prompt before calling generate_media. For a CONSISTENT voice across many clips, bind the chosen id to a saved character (pass it as voice_id in manage_library(kind='character', action='create', ...)) so every generation that reuses the character inherits the same voice. Static list, safe to call without consuming credits.

Inputs: no named parameters. Returns: object

list_music_personasA
Read-onlyIdempotent
Inspect

Scope: read:generations. List the user's saved music personas (newest first). Each entry includes persona_id, name, description, and the source song. Pass a persona_id to generate_music(persona_id=...) to reuse that voice/style.

Inputs: no named parameters. Returns: object

manage_libraryA
Destructive
Inspect

Scope: generate:media. Unified write surface for the user's library. Single tool that adds / renames / deletes both Asset Library entries (images, videos, audio in R2) and Gemini Omni Character entries (reference image + identity description). Reads stay on list_assets and list_gemini_omni_characters — call those first to confirm what exists.

ARGS: action (required) — assets take add | rename | delete; characters take create | delete. Asset args (used when kind='asset'): name — handle for 'add' (lowercase letters/digits/_/-, max 64 chars, unique per user). url — remote source for 'add' (server fetches; validates MIME type: jpeg / png / webp / mp4 / mov / mkv / webm / mp3 / wav / m4a / ogg / weba; 50 MB total + 500 MB-per-file quota). shared — WHERE an 'add' lands for accounts with an active team: true = the shared team library, false = the personal library. This is the USER's choice, never yours: if the account has a team (the list_assets response carries a team block), ASK the user 'personal or team library?' BEFORE saving and pass their answer. Omitting shared on a team account returns shared_scope_required (with team_name) instead of saving — ask, then retry. Accounts with NO team: don't ask and don't pass it; the asset is always personal. asset_id — UUID from list_assets for 'rename' or 'delete'. new_name — new handle for 'rename' (same rules as name). Character args (used when kind='character'): descriptions — appearance / identity / style text for 'create'. image_url — publicly accessible jpeg / png / webp for 'create'. character_name — optional display name for 'create'. voice_id — optional voice from list_gemini_omni_voices for 'create'. character_id — UUID id field from list_gemini_omni_characters for 'delete' (NOT the provider's opaque character_id field). Kling 4.0 library args (kind='kling_element' or 'kling_voice', actions create | delete). Saved entries show in the Kubeez studio's Kling 4.0 'Elements and voices' editor and come back from list_assets under kling_4_elements / kling_4_voices, ready to pass to generate_media's elements / voices on kling-4-0-* models (mention each as @name; Kling 4.0 Flash, kling-4-0-flash-*, takes saved elements but not saved voices): name (required): 1-20 letters (no digits or underscores), unique across the user's Kling 4.0 elements AND voices; identifies the entry for 'delete'. kling_element 'create': EITHER image_urls (2-4 public image URLs, the first is the main view: front face or main angle) OR video_url (one 3-8s single-shot character clip, MP4/MOV, uploaded to Kubeez via get_upload_url); optional tag (character, animal, prop, costume, scene, effect, others; a video element is always character), optional descriptions (max 100 chars), optional voice_url (reference audio bound to a character element). Max 7 saved elements. kling_voice 'create': audio_url (public URL of a clear 5-30s voice recording; an element's voice_url follows the same 5-30s rule). Max 3 saved voices.

USE WHEN: stashing a recurring brand asset, freeing library quota, registering a new Gemini Omni character, or cleaning up. Creating a character (kind='character', action='create') is the FIRST step for any consistent-video request: a recurring spokesperson / mascot, a video series, or a marketing ad set that must keep the SAME person + voice across clips. Bind a voice_id so the voice is locked too, then reuse the saved character_id via character_ids / voice_id on each gemini-omni-video-* / gemini-omni-flash-* generate_media call (see kubeez://skills/model-selection). For destructive actions (delete), CONFIRM with the user first if they didn't explicitly ask — both deletes are irreversible. Past generations that already used the asset URL or character_id keep their outputs; only the library row (and R2 blob, for assets) disappears.

ERRORS: invalid_kind, invalid_action, missing_arg (per action), asset-specific: quota_exceeded, name_taken, unsupported_type, fetch_failed, shared_scope_required (team account + no shared — ask the user personal vs team, retry with their pick); character-specific: unsupported_image_type, invalid_voice_id.

Inputs: kind (required), action (required), tag, url, +13 more. Returns: object

open_pricingA
Read-onlyIdempotent
Inspect

Scope: none. Deep-link to the Kubeez pricing / credit top-up page, with the user's CURRENT balance surfaced in the summary so the agent can phrase the suggestion without an extra get_balance call. Use when a generation fails with insufficient_credits, when estimate shows a cost > balance, or when the user explicitly asks to buy more. Returns { url, summary, current_balance }. Phrase to the user e.g. 'You have N credits, this run needs M — top up at '.

Inputs: no named parameters. Returns: object

open_studioA
Read-onlyIdempotent
Inspect

Scope: none. Return a deep-link the user can open to land in the Kubeez studio with their context prefilled. Use when the agent has done the planning + first generation and the user wants to continue iterating in the UI (richer controls, side-by-side compare, drag-and-drop refs). Optional: prompt, model, aspect_ratio — get appended as query params. Returns { url, summary }. Just hand the url to the user; do NOT try to render the page.

Inputs: model, prompt, aspect_ratio. Returns: object

pick_modelA
Read-onlyIdempotent
Inspect

Scope: none (any authenticated key). Resolve ONE model for a job, with priced alternates and the reason anything was excluded. PREFER THIS over reading the whole get_models catalog when you are choosing a model: the video catalog alone is 141 rows, and comparing them by eye is how a job lands on a model that costs several times what an equally capable one does. Describe the job, not a model: modality, how many images / videos / audios you will attach, duration_s, spoken_language, wants_recurring_character, nsfw, aspect_ratio. DESCRIBE THE INPUTS HONESTLY - pass images=1 when the user attached one image. A wrong count returns a model that cannot take the file, and the attachment gets ignored. Returns primary (the QUALITY pick for that lane), alternates led by the CHEAPEST capable model with its price, rejected naming what could not do the job and why, and cheaper_option when the cheap route is materially less than the primary - say BOTH numbers in your plan card when it is. tier='cheapest' inverts it and returns the cheapest capable model as primary; tier='quality' is the default behaviour stated explicitly. If the user NAMED a model, pass it as user_named_model: it is returned verbatim with its price and a capability_warning if the inputs do not fit. This tool NEVER substitutes a different family or tier for a model the user asked for. A spoken name ('treblo', 'kling 4 flash') that matches exactly ONE id resolves to it and says so in resolved_from; one that matches several returns did_you_mean, so ask. This does not generate anything and costs no credits; follow it with estimate and generate_media.

Inputs: nsfw, tier, audios, images, +8 more. Returns: object

planA
Read-onlyIdempotent
Inspect

Scope: none (any authenticated key). Read-only planner. Input: free-text user intent (REQUIRED — e.g. 'a hero image for my coffee brand', 'animate my product photo', 'add a cigar to the monkey in this video', 'separate vocals from this song'). Optional flags — PASS THESE when relevant or the planner will mis-route: has_local_files=true if the user attached image / audio reference media; has_video_reference=true if the user attached a VIDEO and wants it modified (video-to-video editing — e.g. "add X to this clip", "restyle this video", "make the subject do Y"). When video_reference is true the planner routes to a video-ref model (Gemini Omni Video HD video-ref) instead of text/image-to-video, which is the correct choice for editing an existing clip. Output: a structured plan with recommended_model, estimated_credits, estimated_seconds, requires_upload, suggested_aspect_ratio, workflow_steps[], and skills_to_fetch[]. Use this as the FIRST call of a session when the user describes a goal but hasn't picked a model. It collapses get_models → get_balance → estimate → human-guidance into a single round-trip and tells you exactly which skills to read before generating. After the plan, fetch each item in skills_to_fetch (via get_skill or resources/read kubeez://skills/<name>), then follow workflow_steps in order. This tool does NOT charge credits and does NOT call any model — it's pure deterministic routing on top of the model catalog + skill index. If the intent is ambiguous, the plan returns clarifying_questions[] instead of a model recommendation; ask the user those, then call plan again.

Inputs: intent (required), target_aspect, has_local_files, has_video_reference. Returns: object

play_gemini_omni_voiceA
Read-onlyIdempotent
Inspect

Scope: none. Return a single voice's preview audio URL so the user (or the agent) can hear how a Gemini Omni voice sounds before using it in a generation. Faster than list_gemini_omni_voices when you already know the id (e.g. user said 'play me Kore' or you just picked one from the catalog). ARGS: voice_id — one of the 30 known ids (e.g. kore, puck, achernar). Unknown ids return invalid_voice_id with a hint to call list_gemini_omni_voices. RETURNS: { voice: {id, label, gender, description, preview_url}, preview_url, message }. The preview_url is a ~7s mp3 hosted on the Kubeez web app — surface it directly to the user so they can play it. USE WHEN: the user asks 'what does sound like', 'play that voice', or you want to confirm a pick before passing it to generate_media.

Inputs: voice_id (required). Returns: object

regenerateAInspect

Scope: generate:media. Re-run a past media generation with optional overrides. Use after critique flags an issue: regenerate(id, prompt='same concept but fix the warped hand'). Original row's prompt / model / aspect_ratio / source_media_urls / duration are reused unless you override them. Overrides: prompt (replaces original), seed (force a new seed; 0 = let provider pick), aspect_ratio, negative_prompt, variant_count (1-4 — fans out N parallel calls and returns N generation_ids). Only works on media_generations rows owned by the caller. Music / separation / dialogue regeneration is not supported via this tool — call the original tool again. Cost is identical to a fresh generate_media — confirm with the user before calling, especially for per-duration video models.

Inputs: id (required), seed, prompt, aspect_ratio, +2 more. Returns: object

remove_backgroundAInspect

PRE-FETCH (mandatory): resources/read kubeez://skills/media-upload AND kubeez://skills/billing-confirmation.

Cut the subject out of an image. The output is a PNG with a REAL transparent background (alpha) — use this instead of prompting an image model for a 'transparent background', which bakes a white or checkerboard fill into the pixels. Ideal for logos, product shots, mascots, and cutouts you will composite over something else.

Scope: generate:media. Billing: flat per image — call estimate(model='remove-background') for the exact price, tell the user, and get confirmation before calling this. Takes NO prompt: it is an operation on an existing image, not a generate_media model.

CANONICAL FLOW: (1) if the user has a local file, call get_upload_url(model_id='remove-background') and send them the link. (2) when they confirm the upload, call get_upload_session(token) to get the public URL. (3) quote the price and ask for confirmation. (4) ONLY THEN call this tool.

Required: image_url — a public HTTPS URL of a .jpg / .jpeg / .png / .webp image (max 5MB; the provider rejects other formats). An already-generated Kubeez image works directly — pass its output URL.

Async: returns {generation_id, status: 'pending'} immediately. Poll get_status(generation_id) until status is 'completed'; the output URL is the cutout PNG. Also available in-app at kubeez.com (image tools → Background Remover).

Inputs: image_url (required). Returns: object

trim_videoAInspect

Scope: generate:media. Cut a window out of any publicly accessible video URL and host the trimmed clip — no manual upload required. The server downloads the source, runs ffmpeg -c copy (lossless stream-copy, no re-encode), uploads the trimmed MP4, and returns a public url you can pass straight into generate_media as a source_media_urls value. ARGS: source_url (required — public video URL); start_s (default 0); end_s (required — must be > start_s; range capped at 60s). USE WHEN: a model needs a short reference clip (Gemini Omni Video expects ≤10s; Kling Motion Control caps at 30s) and the user's source is longer than the cap. WHICH WINDOW IS THE USER'S CALL: when the source exceeds the cap, ASK the user which segment to keep (in the Kubeez chat UI, emit the trim_request block instead) and only then trim. Never silently pick a window yourself — agents defaulting to the opening seconds of a long clip discard the moment the user actually wanted. Auto-trim ONLY when the user already named the moment ('use the part at 1:20') or explicitly told you to choose. RETURNS: { url, size_bytes, duration_s, start_s, end_s, elapsed_s }. ERRORS: invalid_trim_range, missing_source_url, source_too_large (provider cap, currently 500 MB), trim_too_long (>60s output), source_fetch_failed (source URL 4xx / unreachable), ffmpeg_failed, processor_timeout, processor_unavailable.

Inputs: end_s (required), source_url (required), start_s. Returns: object

view_mediaA
Read-onlyIdempotent
Inspect

Scope: read:generations. Return one of YOUR generated outputs as inline media, so it is displayed rather than linked. Pass output_id — the id inside outputs[] from get_status / get_generation / get_generations (the OUTPUT id, not the generation id). USE THIS when you want the user to actually SEE an image you just made, and ALWAYS instead of downloading media.kubeez.com yourself: many hosts block that origin in their sandbox, and fetching it externally wastes a turn and a second copy of the file. Images and audio come back as inline media. Video, anything over the 3.5 MiB inline limit, and any CDN failure come back as a short text pointer to the URL instead — MCP has no inline video block. Owner-only: another user's output is refused.

Inputs: output_id (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    Try in Browser

    Glama MCP Gateway

    Add one secure layer between your agents and this server.

    Resources