Connect your app to Replicate

Replicate lets you run machine learning models through one account. Connect it to Mythos when your app needs a particular model for generating or transforming images, video, audio, or text.

Replicate is an App + chat connector: use it in project chat while building and in the app you publish. The models you can run depend on what your Replicate account can access.

Your app can use Replicate to generate images from prompts, transform existing images, create video clips, transcribe recordings, or run a model you have trained. The input and output differ between models, so include the model’s Replicate link when you already know which one you want.

Examples to build

App idea	Prompt to try	What it adds
Image generator	“Build a prompt-to-image tool using this Replicate model: [model link]. Save the generated images in my app.”	A prompt form, generation status, and a gallery of saved results.
Product video maker	“Use a Replicate image-to-video model to turn a product photo into a short clip.”	Let people submit a source image and play the resulting video.
Transcription tool	“Build an audio transcription page with a Whisper model on Replicate.”	Turn recordings into readable transcripts.
Background removal	“Use this Replicate background-removal model for product images.”	Return a processed image and let the user download it.
Personal avatars	“Create a tool that makes illustrated avatar variations from a photo using Replicate.”	Offer a choice of styles and display the generated variations.
Custom model interface	“Build an interface for our model at [Replicate model link], with fields for its inputs.”	Give an existing model a focused interface for your users.

Ask Mythos to check the model’s required inputs before building the form. For example, a video model may require an image URL, while an image model may require a prompt and an aspect ratio.

Connect Replicate to Mythos

You need permission to create App + chat connections in your workspace. An existing saved connection can be linked to another project if you have access to it.

Before you start

You will need a Replicate account and API token. Add billing or credits in Replicate before running a model that requires payment.

Replicate charges for predictions and other paid work under its own pricing. Model runs, training, and deployments can have different costs. Connecting an account does not run a model or include a generation allowance.

1. Create an API token

Sign in to Replicate.
Open Account settings → API tokens.
Create a token named for its purpose, such as Mythos media.
Copy the token and store it securely. Tokens start with r8_.

Use the token in the setup form, never in a chat prompt or public app file. See Replicate’s API token guide for account-side controls.

2. Save the connection

Open Connectors from the Dashboard and select Replicate. If a connection already exists, select Add connection to open a new form.
Enter a Display name, such as Product media.
Paste the token into API token.
Under Sharing, leave the connection Private or use Share with others to select members. Choose Share with workspace if the whole workspace should use it.
Click Connect Replicate. Mythos checks your account without starting a prediction.

In your project, open Chat actions (+) → Project connectors, select the saved Replicate connection, and click Use in this project.

A useful first request is:

Show me the inputs expected by this Replicate model: [model link]. Then build a form for those inputs, with a visible generation status and saved output.

Running a model spends Replicate credits and can proceed without a separate approval question in chat. Review the model and inputs in your request before starting the first run. When a prediction is still running, check its status instead of submitting the same inputs again. Training and deployment changes follow the connection’s approval setting.

Limitations

Output links are temporary. Replicate-hosted output can expire. Ask Mythos to save images, audio, and video in your app’s storage when it builds the feature, rather than displaying only a temporary URL. See Replicate’s output-file guidance.
Tokens are not rotated automatically. When a token is replaced, create a new connection with that token.
One connection uses one account. Your app’s visitors do not each authorize their own Replicate account through this connection.
A model needs its own valid inputs. A working account connection does not guarantee that a particular model version or input combination will run.

If a prediction fails

An Insufficient credit response means the Replicate account needs billing or credits. Check Replicate billing even if the connection itself still appears connected.

For an input error, ask Mythos to compare the submitted values with the chosen model’s input requirements. If the result is pending or unclear, look up the existing prediction before running it again. If a completed result no longer loads, check whether its temporary file URL has expired.

Manage the connection

Open Connectors → Replicate and select the connection. Use Unlink projects for selected projects, or Disconnect to stop access through the saved connection.

Disconnecting does not cancel a prediction, training job, or deployment already running in Replicate. Review those resources in Replicate separately. Saved files in your app are not removed by disconnecting. To change the token, disconnect the old connection, create a new one with the replacement token, and link it to the affected projects. Ask Mythos to set up their app features again.