An example Cog model that demonstrates cog.Secret inputs by calling the Replicate API from inside a prediction.

Given an input image, the model:

1. Sends the image to anthropic/claude-4-sonnet to generate a detailed prompt describing it.
2. Feeds that prompt to black-forest-labs/flux-dev to re-create the image.
3. Returns the generated image.

The Replicate API token is declared as a cog.Secret input:

from cog import Input, Secret

def run(
    replicate_api_token: Secret = Input(
        description="Replicate API token used to call other models",
    ),
) -> Path:
    client = Client(api_token=replicate_api_token.get_secret_value())
    ...

cog.Secret redacts its value in logs and string representations. Read the underlying value with get_secret_value().

Avoid passing the token literally on the command line, since it can leak through your shell history and process listings. Instead, read it from an environment variable:

export REPLICATE_API_TOKEN=r8_...   # set once, ideally via a secrets manager / not inline in shared shells
cog predict -i image=@cat.png -i replicate_api_token="$REPLICATE_API_TOKEN"

You can also read the token from a file (for example -i replicate_api_token="$(cat token.txt)") if that fits your workflow better.

Note: cog.Secret redacts the value in model logs and string representations, but it cannot protect a secret that is already exposed by your own shell history, environment, or process listing. Keeping the token out of those places is your responsibility.