pip install replicate

Start by setting a REPLICATE_API_TOKEN environment variable in your environment. You can create a token at replicate.com/account/api-tokens.

Then use this code to initialize a client:

import replicate

That's it! You can now use the client to make API calls.

If you want to explicitly pass the token when creating a client, you can do so like this:

import os
import replicate

client = replicate.Replicate(
    bearer_token=os.environ["REPLICATE_API_TOKEN"]
)

Create a reference to a model that can be used to make predictions.

import replicate

claude = replicate.use("anthropic/claude-sonnet-4")

output = claude(prompt="Hello, world!")
print(output)

banana = replicate.use("google/nano-banana")
output = banana(prompt="Make me a sandwich")
print(output)

Note: The replicate.use() method only returns output. If you need access to more metadata like prediction ID, status, metrics, or input values, use replicate.predictions.create() instead.

Run a model and wait for the output. This is a convenience method that creates a prediction and waits for it to complete.

import replicate

# Run a model and get the output directly
output = replicate.run(
    "anthropic/claude-sonnet-4",
    input={"prompt": "Hello, world!"}
)
print(output)

Note: The replicate.run() method only returns output. If you need access to more metadata like prediction ID, status, metrics, or input values, use replicate.predictions.create() instead.

For cases where you need to make direct API calls not covered by the SDK methods, you can use the low-level request interface:

import replicate

client = replicate.Replicate()

# Make a custom GET request
response = client.get("/custom/endpoint")

# Make a custom POST request with data
response = client.post(
    "/custom/endpoint",
    json={"key": "value"}
)

# Make a custom request with all options
response = client.request(
    method="PATCH",
    url="/custom/endpoint",
    json={"key": "value"},
    headers={"X-Custom-Header": "value"}
)

See the README for more details about response handing, error handling, pagination, async support, and more.