Replicate
Run models on Replicate, poll or wait for their output, and manage deployments and trainings.
About
Replicate ships in the w6w first-party pack. It declares 20 actions, 2 health checks, and the host runs its code in a sandbox that never sees the credential.
- App id
io.w6w.replicate- Version
- 0.1.0
- Author
- w6w
- Licence
- MIT
- Categories
- AI & Machine Learning · Developer Tools
Build with Replicate
Three routes to the same 20 actions. The Workflow tab is generated from Replicate's own manifest and carries its real ids, so it is copy-pasteable; the Code and CLI examples are the same call for any action on any app, so every app-specific value in them is a blank you fill in.
Run a deployment
deployment-prediction-create Start a prediction on a deployment — a model with warm hardware behind it.
Get a model's readme
model-readme-get The author's documentation, as Markdown — what the input fields actually mean.
Get a model version
model-version-get One version and its input schema — what a prediction on it must send.
List a model's versions
model-version-list A model's versions, newest first. Official models have none.
Run a model version
prediction-create Start a prediction on a pinned model version. Returns before the model has run.
Run a model
prediction-create-from-model Start a prediction on a model's current version. Returns before the model has run.
Cancel a training
training-cancel Stop a training, and its billing — which for a training is worth real money.
Start a training
training-create Fine-tune a model into a destination you already own. Runs for minutes to hours.
A workflow step names the app and the action, and the editor fills in the
connection when you pick one. This is the Step shape from the
workflow spec, carrying Replicate's real ids.
{
"manifestVersion": "2",
"name": "replicate-example",
"steps": [
{
"id": "deployment-prediction-create",
"uses": {
"app": "io.w6w.replicate",
"action": "deployment-prediction-create",
"connection": "conn_YOUR_CONNECTION_ID"
},
"with": {
"deployment": "<deployment>",
"input": "<input>"
}
}
]
}Here are some of the things you can do
-
Run a deployment
performdeployment-prediction-create -
Get the account
readaccount-get -
List collections
readcollection-list -
List deployments
readdeployment-list -
List hardware
readhardware-list
+15 more actions available
Every app-specific value here is a blank you have to fill in. An
app action is reached through the connection that authenticates it, so the
address is a connection id, not the app id — and connections belong to your account,
so a public page cannot know yours. Create one for Replicate, then fill in
the three blanks: conn_YOUR_CONNECTION_ID, the action key, and the
parameters that action declares. The call itself is real — the shape is transcribed
from the studio's own snippet builder, which prints the same kind of blanks — but
nothing in it is specific to Replicate. The Workflow tab is where this app's
real ids are.
npm install @w6w/sdkyarn add @w6w/sdkpnpm add @w6w/sdkdeno add npm:@w6w/sdkimport { W6wClient, isActionRun } from "@w6w/sdk";
// Reads W6W_BASE_URL and W6W_TOKEN from the environment when omitted.
const client = new W6wClient();
const envelope = await client.run({
urn: "conn_YOUR_CONNECTION_ID",
action: "deployment-prediction-create",
payload: {
deployment: "<value>",
input: "<value>",
// waitSeconds: "<value>",
// webhook: "<value>",
// webhookEventsFilter: "<value>",
},
});
if (isActionRun(envelope)) console.log(envelope.value);npm install -g @w6w/cliw6w run conn_YOUR_CONNECTION_ID --action deployment-prediction-create --payload '{"deployment":"<value>","input":"<value>"}'Give an AI agent Replicate — without giving it Replicate's credentials. One MCP endpoint exposes every app, function and workflow the caller is entitled to, as tools it can discover and run. Access is granted per team while we onboard.
One tool call{
"name": "w6w_invoke",
"arguments": {
"ref": "app:io.w6w.replicate#deployment-prediction-create",
"input": {
"deployment": "<deployment>",
"input": "<input>"
}
}
}
Every tool names its target with a single ref. The
app: form above doesn't name a connection at all — the
host resolves which of the caller's Replicate connections to sign
with, and refuses rather than guesses when the answer is ambiguous.
What the agent gets
Credentials it can't read
The token is attached host-side, at the moment of the call. It is never a tool argument, never in the model's context, and never in a transcript — so a prompt injection has nothing to exfiltrate.
A tool surface scoped to the caller
Tools are derived per end user from what that person has actually connected and is entitled to — not one shared bot identity carrying the union of everyone's access.
A durable workflow in one call
Multi-step work runs on the workflow engine and returns a run handle the agent can poll — retries, branching and state survive the conversation that started them.
Health-aware discovery
Replicate's declared health checks are on the surface too, so an agent can tell "the vendor is down" from "your credential expired" before it burns a retry on either.
The MCP surface is part of the hosted platform. Replicate itself is MIT, and the runtime that executes it is source-available (FSL).
Health checks
Replicate declares its own checks, so its health is a property of the app rather than something the host guesses at.
Replicate platform status
The API and prediction components on Replicate's own status page. Unauthenticated and unsigned.