API and MCP · developers
Talking avatar video from code
Percify's AI avatar API turns an image, a voice and a script into a lip synced talking video through one REST endpoint: send a model and its inputs to /run, then poll the generation until the file is ready. The same models are reachable from AI agents through an MCP server. API keys come with every account and draw on the same credit meter as the app, or pay as you go with API funds from $10 at $0.02 a credit, no plan needed. It renders video files; it is not a realtime streaming avatar.
- 1
- REST endpoint to run any model
- 120+
- image, video and voice models
- 65
- credits for the first agent made 11 s video
- $0.02
- a credit, pay as you go from $10
Credits cost about 2 cents each depending on the plan, so a minute of talking video at 120 credits works out at roughly $2 to $3.
How a talking avatar is built through the API
Three calls, each the input of the next: an image model makes or edits the face, a voice model speaks the script, and a lip sync model drives the mouth from that audio. The first full agent session ran exactly this pipeline and finished an 11 second talking video for 65 credits.
| Step | Model type | Example |
|---|---|---|
| Face | Image | GPT Image 2, or your own photo |
| Voice | Text to speech or voice clone | Zonos 2, which clones from a reference recording |
| Video | Lip sync | InfiniteTalk, in a fast and a standard variant, billed per second |
The endpoints
The base URL is https://api.percify.io/v3/playground/v1, and every request carries your pk_live key as a Bearer token in the Authorization header. The replication pipeline is exposed as well: send a short video URL to be analysed and read back the shot by shot blueprint.
| Endpoint | What it does |
|---|---|
| POST /run | Start a generation with a model id and its inputs |
| GET /generations/{id} | Status and the output URL |
| GET /models | Every model and the inputs it accepts |
| GET /openapi.json | The machine readable spec |
For AI agents: the MCP server
The MCP server lives at https://mcp.percify.io/mcp. In Claude, add it as a custom connector and sign in with your Percify account, no key needed. In Claude Code, Cursor, Codex or any other MCP client, pass your pk_live key as a Bearer header. An agent can then make the face, the voice and the video from inside the chat. Setup is on the MCP page.
Timing, and designing for long jobs
Lip sync on the fast 480p path spends about 7.62 seconds of compute per second of audio, and a four minute clip on the fast model completed in about 13.6 minutes. Treat every render as an asynchronous job: start it, store the id, and poll with a backoff rather than holding a request open.
That also marks the edge of what it is for. It produces finished video files. If you need an avatar that answers people live in a conversation, this is the wrong tool.
What it costs
The API uses the same credit meter as the app, with lip sync billed per second of output. A 20 second talking clip at 480p is 40 credits on the fast model (2 a second). Without a plan, add API funds on dash.percify.io from $10 at $0.02 a credit; with one, runs use plan credits first. Current rates are on API pricing, and the model list with every input is in the AI model catalog.
Your first render in four steps
- 1
Get a key
API keys come with every plan. Keep the pk_live key server side.
- 2
List the models
GET /models returns every model and the inputs it accepts.
- 3
Run the pipeline
Face, then voice, then lip sync, each a POST /run with the previous output as input.
- 4
Poll and store
Read /generations/{id} until the output URL appears, then keep your own copy.
Output from the same pipeline
Real talking avatar renders, the kind the API returns.
Alex
talks from one photo
Christine
a spokesperson close up
News anchor
seated, steady framing
Frequently asked questions
Does Percify have an AI avatar API?
Yes. One REST endpoint runs any model, and a talking avatar is three calls: an image, a voice, then lip sync. The same models are available to AI agents through an MCP server.
Which plan includes API access?
None is needed. Every account can create API keys. Runs use plan credits first, then API funds, which you add pay-as-you-go on dash.percify.io from $10 at $0.02 a credit.
How long does an API render take?
Lip sync on the fast 480p path takes about 7.6 seconds of compute per second of audio. A four minute clip finished in about 13.6 minutes, so design for asynchronous jobs.
Can AI agents like Claude use it?
Yes. Connect the MCP server at https://mcp.percify.io/mcp as a custom connector in Claude and sign in, or pass a pk_live key in Claude Code, Cursor or Codex.
Is it a realtime streaming avatar API?
No. It renders finished video files. It is not built for an avatar that holds a live conversation.
Keep reading
Build it into your product
One key, one endpoint, the whole avatar pipeline.