Replicate
Connect Replicate to find and run AI models for images, video, audio and text, and manage your models, trainings and deployments.
You paste one value into PopMCP: a Replicate API token, which starts with r8_.
- Setup time
- 1 min
- Auth
- API credentials
- API base URL
- https://api.replicate.com/v1
- Rate limits
- Replicate allows 600 run creations per minute and 3000 requests per minute on everything else, with short bursts above that. Limits get stronger when credit runs low, and accounts with granted credit and no payment method get 1 request per second, up to 6 per minute. Over the limit it answers 429 and says when the limit resets.
Before you begin
- A Replicate account. You sign in to Replicate with a GitHub account
- Prepaid credit on the account, if your AI will run models. Replicate takes payment up front under Account, then Billing. Searching models and reading past runs need no credit
Get your credentials
Open your API tokens in Replicate
Sign in to Replicate and open replicate.com/account/api-tokens. To connect an organization, use a token that belongs to that organization. A token reaches everything on its account and has no permission options, so choose the right account now. You can connect another later with its own token.
Create a token for PopMCP and copy it
Replicate makes a default token when you sign up, and you could use that one. A separate token is better: give it a name like PopMCP and create it, then copy it. You can later turn this one off without touching your other apps.
API tokens
Tokens authenticate your requests to Replicate's API.
This view is a simplified sketch: the labels on your page may differ slightly. Tokens are 40 characters long and start with r8_.
Paste the token into PopMCP
In PopMCP, paste the token into API token and connect. PopMCP checks it by reading your account name from Replicate. New connections start read-only, so your AI can search models and read past runs but cannot start a run or spend credit until an owner or admin allows write actions.
Connect in PopMCP
Open your workspace, choose Replicate in the marketplace, and fill in the connect dialog:
| Field | What to enter |
|---|---|
| API tokensecret | In Replicate open Account, then API tokens, and copy a token. It starts with r8_. Stored encrypted and used only on our servers.Encrypted with AES-256-GCM on save and never shown again. |
PopMCP verifies the credential against Replicate immediately. On success the connection flips to Verified and its MCP tools are provisioned.
Verify it worked
PopMCP marks the Replicate connection Verified and labels it with your account name and username. Asking the AI to find a model for a task, then to show that model's inputs, returns real results at no cost.
Troubleshooting
"Replicate did not accept this API token..."
Why: The token was copied incompletely or has been turned off.
Fix: Open replicate.com/account/api-tokens and copy a token that is still active. It starts with r8_ and is 40 characters long.
"That is a webhook signing secret..."
Why: A value starting with whsec_ was pasted. It proves webhooks came from Replicate and cannot call the API.
Fix: Paste an API token from replicate.com/account/api-tokens instead.
The AI says it cannot run a model
Why: The connection is read-only, which is how every new connection starts. Running a model spends credit, so it counts as a write.
Fix: Have an owner or admin allow write actions on the connection in PopMCP.
Runs are refused because of credit, or stop starting
Why: The account's prepaid credit is used up. At zero, Replicate starts no new work and shuts down what it is running for you.
Fix: Add credit under Account, then Billing, in Replicate. Auto reload keeps it from running out.
"Nothing was started and nothing was spent..."
Why: An input does not match the model: an unknown name, a value outside its limits, or a file given as a path on a computer.
Fix: The message names each problem. The AI can read the model's inputs and try again. For a file with no public link, it can upload the file first.
An output link no longer opens
Why: Replicate removes the inputs, outputs, files and logs of an API run about an hour after it runs.
Fix: Save what you want to keep as soon as the run ends. Past that, run the model again.
A run says it is still going when the AI stops waiting
Why: Long models outlast one wait. The run carries on at Replicate, and it stops by itself after 30 minutes.
Fix: Ask the AI to keep waiting for the same run. It has the id. Starting again would pay twice.
Tools fail with 429
Why: Replicate allows 600 run creations and 3000 other requests per minute. Limits are stronger when credit runs low, and only 6 requests per minute with granted credit and no payment method.
Fix: The answer says when the limit resets. Adding a payment method or more credit lifts the lower limits.