Your media stops living on someone else's CDN
On v2 you sign a PUT per job and the render is written into your bucket. We keep no copy, so there is no expiring link to babysit and no third-party host in your delivery path.
The video API that does not reject your traffic.
Image-to-video over REST, no arbitrary content filters, and every render written straight into your own bucket.
# one call, one clip curl -X POST \ https://api.musebox.ai/api/v1/jobs \ -H "Authorization: Bearer $MBX_KEY" \ -H "Content-Type: application/json" \ -d '{ "type": "i2v", "modelId": "tmpl-i2v-classic", "inputUrl": "https://cdn.you/frame.jpg", "quality": "fast" }'
status: done
550+templates in the shared catalogue
4job types: i2v, extend, t2i, audio
1020pxtop quality tier, longest side
HMACsigned webhooks with retries
USDTbilling rail for partners
Three reasons teams migrate to us
Every partner who moved here was blocked by one of these somewhere else.
Your prompts stop coming back rejected
General-purpose video APIs reject the work your product exists to do. We do not apply those creative filters. Illegal content stays prohibited and every account is bound by our terms, but the model does not second-guess your catalogue.
Your characters become templates
We train custom LoRA models on your footage and expose them as private templates that only your keys can call.
Already carrying production traffic
This is not a beta waiting for its first integration. Partner accounts run through the same queue and the same GPU fleet every day.
Millions
generations delivered through the API every day
99.9%
job success rate, the same number your console reports
The traffic comes from three kinds of product, and it is concentrated in the United States, Europe and Brazil.
Mobile apps
Telegram bots
Established AI generators
Submit, render, deliver
The whole integration is one POST and one webhook handler. Everything else is optional.
Submit
POST /v1/jobs with a template id and a public image URL. You get a job id, a queue position, the credits charged and your balance after, all in the same response.
Render
The job runs on dedicated GPU workers. Text-to-image lands at roughly 60 seconds p50 and 150 seconds p95 when there is queue depth. Poll if you have to, but webhooks are the intended path.
Deliver
We sign the callback with your per-key secret and retry on failure. On v2 the file is already sitting in your bucket by the time the webhook arrives.
Real output from the shared catalogue
Each clip below is one call against a template every partner account can reach on day one.
What your users end up looking at
You own the interface. One account and one key drive a storefront and a chat bot alike, because both are the same two calls underneath.
550+ templates from /v1/models
Your server posted the job and is waiting on the webhook.
Delivered into your bucket, served from your own CDN.
Save to gallery
A storefront in your app
Your catalogue, your pricing, your credits. The API supplies the templates and the render.
Send a photo and pick a style.
Queued. This takes about 40 seconds.
A bot in a chat
One POST when the photo arrives, one webhook when the clip is ready. No polling loop to run.
The render never touches our CDN
On v1 the output sits in our storage behind a link that expires about 24 hours after the job finishes. On v2 you attach a presigned PUT to the job and the worker writes the finished file directly into your bucket. We keep nothing.
AWS S3
Cloudflare R2
Google Cloud Storage
Azure Blob
Backblaze B2
S3-compatible
// add one block to any v2 job { "type": "i2v", "modelId": "tmpl-i2v-classic", "inputUrl": "https://your-cdn.com/frame.jpg", "delivery": { "url": "https://bucket.r2.cloudflarestorage.com/out.mp4?X-Amz-...", "method": "PUT", "headers": { "Content-Type": "video/mp4" } } } // the callback is a delivery receipt, not a download link { "jobId": "...", "delivery": { "status": "delivered" } }
Working code in three languages
Create a job, then verify the callback signature before you trust it. That is the entire contract.
curl -X POST https://api.musebox.ai/api/v1/jobs \ -H "Authorization: Bearer $MBX_KEY" \ -H "Content-Type: application/json" \ -d '{"type":"i2v","modelId":"tmpl-i2v-classic","inputUrl":"https://your-cdn.com/frame.jpg","quality":"fast"}' # poll only if you cannot take a webhook curl https://api.musebox.ai/api/v1/jobs/$JOB_ID \ -H "Authorization: Bearer $MBX_KEY"
import crypto from 'node:crypto'; const res = await fetch('https://api.musebox.ai/api/v1/jobs', { method: 'POST', headers: { Authorization: `Bearer ${process.env.MBX_KEY}`, 'Content-Type': 'application/json', }, body: JSON.stringify({ type: 'i2v', modelId: 'tmpl-i2v-classic', inputUrl: 'https://your-cdn.com/frame.jpg', quality: 'fast', webhookUrl: 'https://you.example/hooks/musebox', }), }); const { data } = await res.json(); // reject anything you cannot verify function verify(rawBody, signature, secret) { const expected = crypto.createHmac('sha256', secret) .update(rawBody).digest('hex'); return crypto.timingSafeEqual( Buffer.from(expected), Buffer.from(signature) ); }
import os, hmac, hashlib, requests r = requests.post( "https://api.musebox.ai/api/v1/jobs", headers={"Authorization": f"Bearer {os.environ['MBX_KEY']}"}, json={ "type": "i2v", "modelId": "tmpl-i2v-classic", "inputUrl": "https://your-cdn.com/frame.jpg", "quality": "fast", "webhookUrl": "https://you.example/hooks/musebox", }, timeout=30, ) job = r.json()["data"] # reject anything you cannot verify def verify(raw_body: bytes, signature: str, secret: str) -> bool: expected = hmac.new(secret.encode(), raw_body, hashlib.sha256).hexdigest() return hmac.compare_digest(expected, signature)
What you can call, and what holds it up
generation
- Image to video
- A still in, a five second clip out.
POST /v1/jobs - Long-form sequences
- Chain up to five models into one continuous 15s+ video.
POST /v1/jobs/sequence - Text to image
- Free-form scene prompt, seeded for reproducibility.
POST /v1/jobs/t2i - Audio track
- Optional add-on on plans that carry the capability.
catalogue
- 550+ shared templates
- Listed through
GET /v1/models, no prompt engineering needed. - Cover tags
- Filter the catalogue by cover look so the artwork fits the audience you serve.
- Private LoRA models
- Trained on your footage, callable only by your keys.
- Four quality tiers
- 720, 832, 960 and 1020px on the longest side, gated per plan.
operations
- Rate limits you can read
- 60 to 1000 requests per minute by plan, with
RateLimit-*headers on every response. - Concurrency by plan
- 5 concurrent jobs on entry plans, up to 100 on Studio.
- Signed callbacks
- Per-key HMAC secret plus an automatic retry schedule.
- Documented failures
- Every error code maps to one action, including the 422 that tells you a presigned URL will not work before you are charged.
Four plans, quoted per project
Plans differ by what you can call, how fast you can call it, and how many credits come with the month. We quote once we know your volume.
Pro
Quoted on request
- Image to video, extend and audio
- 832p and 960p quality tiers
- 200 requests per minute
- 25 concurrent jobs
- Delivery into your own bucket
Pilot
Quoted on request
- 30 day evaluation window
- Image to video and extend
- 720p, 60 requests per minute
Starter
Quoted on request
- Image to video at 720p
- 60 requests per minute
- Production volume allowance
Studio
Quoted on request
- 1000 requests per minute
- 100 concurrent jobs
- Custom SLA and dedicated capacity
Credits never expire while the subscription is active, and stay usable for 30 days after a cancellation. Requests stop at a zero balance unless your plan carries overage, and quota warnings arrive at 80, 95 and 100 percent.
Built for the teams nobody else will serve
Adult platforms
The catalogue matches what your audience already pays for, the renders land in your storage rather than ours, and USDT settlement keeps billing working when card processors will not.
Telegram bots
One POST per user request and a webhook that fires when the file is ready. No long-lived connections and no polling loop to babysit while your bot serves thousands of chats.
Agencies and white-label resellers
Resell generation under your own brand. Cover tags let you present a catalogue that fits each regional audience, and private LoRA models keep your clients' characters yours.
Companion and character apps
Animate the same character consistently across a whole product. Train a private model once, then call it like any other template for as long as the account lives.
Questions partners ask first
What content does the API allow?
We do not apply the arbitrary creative filters that general-purpose video APIs use, which is the reason most partners move here. Illegal content is strictly prohibited, all use is bound by our Terms of Service, and we run a content removal process for reports.
Where do the finished renders end up?
On v1 the render lands in our storage and you download it from a link that expires about 24 hours after the job completes. On v2 you supply a presigned PUT per job, the worker writes the file straight into your bucket, and we keep no copy at all.
Which storage providers does delivery support?
Anything that accepts a presigned PUT: AWS S3, Cloudflare R2, Google Cloud Storage, Azure Blob, Backblaze B2, and generic S3-compatible endpoints. The documentation carries a working presign recipe for each one, plus a self-test endpoint that proves your URL works before you spend credits on a real job.
How long does a generation take?
Text-to-image runs at roughly 60 seconds p50 and 150 seconds p95 once there is queue depth. The first request after an idle period is slower while the model loads, then subsequent jobs run hot.
How do I get an API key?
Keys are provisioned by hand after a short review of your project. Send an access request describing what you are building and your expected volume, and you receive a live key plus a webhook signing secret. There is no self-service signup, which is deliberate.
How does billing work?
Plans are billed monthly and carry a credit allowance. Credits never expire while your subscription is active and remain usable for 30 days if you cancel. Requests hard-stop at a zero balance unless your plan includes overage.
Can I bring my own characters?
Yes. We train custom LoRA models on your reference footage and expose them in your catalogue as private templates that only your keys can call.
Do you accept crypto?
Yes. USDT on TRC20 is the default rail for international partners, which matters for teams that traditional processors will not serve.
Tell us what you are building
Send your project, your platform and your expected volume. We review access requests by hand and reply with a quote and a key.