Skip to main content

1Connect

Add onzero as an MCP server in your client — Claude, Codex, Cursor or any other:

Server URL
https://onzero.ai/mcp
Auth — personal Bearer token (Settings → MCP), works in every client. For claude.ai — OAuth, no token:Authorization: Bearer oz_xxxxxxxxxxxxxxxxFrom the moment you connect, your client gets 67 tools + the professional workflow rules — nothing to explain.
Add it in your client

One command in your terminal:

claude mcp add --transport http onzero https://onzero.ai/mcp --header "Authorization: Bearer oz_..."

2The Claude Skill

The connection gives Claude the tools. The skill gives it the judgment in between.

The skill adds no capabilities — same tools, same models. What it changes is how many attempts it takes to get there: what to ask before generating, how to plan a multi-step production, when to fix a result instead of shipping it. Fewer regenerations, fewer wasted credits.

Download the skillZIP · 12KB · free
Claude CodeUnzip into your skills directory — personal (all projects) or scoped to one project. Restart Claude afterwards.
claude.ai / DesktopSettings → Capabilities → Skills → Upload skill, and pick the ZIP as downloaded — no need to unzip it.
Codex & other clientsNo skill mechanism — paste the contents of SKILL.md into your project instructions file (AGENTS.md or equivalent).

Check it works: ask for "an Instagram post and a story for my brand". Working — Claude asks about the brand, the exact copy and the platform before generating, then produces one master and adapts it to the second size. Not working — it generates two separate images straight away.

This is a local copy that does not update itself. The skill tells Claude to call list_video_models and friends for the live list, so it degrades gracefully — but re-download when new capabilities ship.

3The fixed workflow

This is what correct work looks like, brief to deliverable.

Character — once, locked forever

Exact artwork goes in as reference via create_character, stored pixel-perfect. Every image and video keeps the identity.

Brand assets

Need a logo that doesn't exist? Claude asks for it, saves via create_asset, and uses it exactly in every future creation.

Master, then adapt

One approved master → adapt_image_size reproduces the exact design at every ratio. Never regenerate per size.

Video — consult before generating

Claude presents 2–3 model + technique options with trade-offs — you choose. Per-model stability scaffolding applies automatically.

Full sound

Voiceover (in the character's voice!), music, songs with lyrics and SFX — then add_audio_to_video scores any clip.

Quality gate

Text verbatim? Identity preserved? No artifacts? Fix until it passes. Never ship "almost right".

4Video models

Always available via list_video_models — with live pricing.

ModelBest forDurationsCapabilitiesCredits / 5s
Veo 3.1Cinematic quality, synced dialogue & SFX4/6/8sAudio End frame34
Seedance 2.0Best identity consistency, multi-shot, mascots4–15s End frame200
Kling v3Expressive character motion, up to 6 shots4–15sAudio End frame57
Hailuo 2 ProPhysics, [Push in] camera control6s End frame64
Runway Gen-4.5Clean single-scene reliability5/10s 80
PixVerse 4.5Stylized / anime, 20+ camera moves5/8s 27
Wan I2VFaithful image animation, budget~5s 54
Luma ModifyRestyle existing footage (V2V) 74

5Sound models

Available via list_audio_models. Character selected? The voiceover speaks in their voice.

ModelCategoryBest for
Gemini VoiceVoiceoverFast, multilingual, character voice
ElevenLabs v3VoiceoverMost natural — ads & trailers
MiniMax SpeechVoiceover300+ voices, 30+ languages
Lyria 2MusicGoogle, 48kHz, negative prompt
Stable Audio 2.5MusicUp to ~3 min in one generation
ACE-StepMusicDrafts & loops — cheapest
MiniMax MusicSongVocal songs — [Verse]/[Chorus]
ElevenLabs SFXSFXText-to-sound-effects

6Advanced techniques

Keyframe interpolationFirst + last frame → endImageUrl — the model animates precisely between them.
Subject LockLocks the design to frame one — no new limbs, no morphing.
Reference-to-VideoUp to 9 reference images as @Image1..@Image9 — total consistency in video.
Long-form contentChained extend_video (4/6/8s) beats any single-model length limit.
Scoringadd_audio_to_video — loop for music beds, fit for voiceover, trim for SFX.
Talking charactersgenerate_voiceover in the character's voice → lipsync_video syncs the lips.