A typed client library for the replicate.com API.
You can use this to access the prediction API in a type-safe and convenient way.
Just install it with your favorite package manager:
yarn add replicate-api
pnpm add replicate-api
npm install replicate-apiThe package should work in the browser and in Node.js versions 18 and up.
You need an API token for nearly all operations. You can find the token in your account settings.
You can create a new prediction using the
stability-ai/stable-diffusion model and wait for the result
with:
const prediction = await predict({
model: "stability-ai/stable-diffusion", // The model name
input: { prompt: "multicolor hyperspace" }, // The model specific input
token: "...", // You need a token from replicate.com
poll: true, // Wait for the model to finish
})
console.log(prediction.output[0])
// https://replicate.com/api/models/stability-ai/stable-diffusion/files/58a1dcfc-3d5d-4297-bac2-5395294fe463/out-0.pngThis does some things for you like resolving the model name to a model version and polling until the prediction is completed.
const result = await predict({ model: "replicate/hello-world", input: { prompt: "..." }, token: "..." })Then you can check result.status to see if it's "starting", "processing" or succeeded. If it's "succeeded" you
can get the outputs with result.outputs. If not you can check back later with getPrediction() and the id from
result (result.id).
You can also set poll: true in the options of predict() to wait until it has finished. If you don't do that, you can
still use .poll() to poll until the prediction is done.
// If you have a PredictionState:
const finishedPrediction = prediction.poll()
// If you only have the prediction ID:
const finishedPrediction = await pollPrediction({ id, token: "..." })
// If you are creating a new prediction anyways:
const finishedPrediction = await predict({ ...otherOptions, poll: true })// If you have a PredictionState:
const currentPrediction = prediction.get()
// If you only have the prediction ID:
const currentPrediction = await getPrediction({ id, token: "..." })// If you have a PredictionState:
const currentPrediction = result.cancel()
// If you only have the prediction ID:
const currentPrediction = await cancelPrediction({ id, token: "..." })Canceling the prediction also returns the state of the prediction after canceling.
const info = await getModel({ model: "replicate/hello-world", token: "..." })const info = await listVersions({ model: "replicate/hello-world", token: "..." })The first example used a few convenience functions to make it easier to use the API. You can also use the lower-level functions that map the API calls more directly.
const model = await getModel({ model: "stability-ai/stable-diffusion", token: "..." })
let prediction = await predict({
version: model.version,
input: { prompt: "multicolor hyperspace" },
token: "...",
})
// pollPrediction does this a bit smarter, with increasing backoff
while (prediction.status === "starting" || prediction.status === "processing") {
await new Promise(resolve => setTimeout(resolve, 1000))
prediction = await getPrediction({ id: prediction.id, token: "..." })
}
console.log(prediction.outputs[0])
// https://replicate.com/api/models/stability-ai/stable-diffusion/files/58a1dcfc-3d5d-4297-bac2-5395294fe463/out-0.pngconst result = await listPredictions({
token: "...",
})Returns up to 100 predictions. To get more, use the next function:
const moreResults = await result.next()You can also set all: true to get all predictions.
To use file inputs you need to pass them as URLs. You can use the loadFile function to convert local files to base64
data URLs:
const testaudioURL = await loadFile("./testaudio.mp3")
//You can also use an HTTPS URL to load files from the web.
You can create a new prediction for the openai/whisper model and wait for the
result with:
const prediction = await predict({
model: "openai/whisper", // The model name
input: {
audio: await loadFile("./testaudio.mp3"), // Load local file as base64 dataurl
// audio: "https://raw.githubusercontent.com/zebreus/replicate-api/master/testaudio.mp3", // Load from a URL
model: "base",
}, // The model specific input
token: "...", // You need a token from replicate.com
poll: true, // Wait for the model to finish
})
console.log(prediction.output.transcription)
// Transcribed text- replicate-js - A js object-oriented client for replicate
This package uses the fetch API which is only supported in Node.js 18 and up. If you need to use an older version of
Node.js, you can use node-fetch. It will be detected and used automatically if your node does not provide a native
fetch. The Options object supports passing a custom fetch function, you can also try to pass node-fetch there.
To run the tests for this package you need an API token from <replicate.com>. Then you create a src/tests/token.ts
file that exports the token as a string like export const token = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx". Now you
can run yarn test to run the tests.
