AI and media

AI agents that use Replicate

UtilitySpaces gives an AI agent scoped access to Replicate through its own encrypted connection, not your password. The agent can call 5 named Replicate tools, and the one that writes waits for a person's approval unless an admin has set the connection to run without asking. Every call is written to a hash-chained audit log with its arguments and its result.

What an agent can do in Replicate

Browse Replicate models and run them as predictions for image, video, and audio generation. An agent calls these by name, and matching is exact.

Every tool a Replicate connection exposes, and whether it writes, which waits for approval by default
Tool	What it does	Waits for approval by default
List models	List public models on Replicate.	No — read only
Get model	Get one model's metadata and its latest version id.	No — read only
Create prediction	Start a prediction (billed run). Predictions are asynchronous — poll get_prediction until status is succeeded or failed.	Yes — it writes
Get prediction	Get a prediction's status and output.	No — read only
List collections	List curated model collections.	No — read only

4 read tools, 1 write tool. Generated from the connector catalogue the product runs on, so it cannot list a tool that does not exist.

What teams hand an agent in Replicate

Teams that generate media hand an agent the model work on Replicate: List models and Get model to find the right one and its version, Create prediction to run it, Get prediction to collect the result. Predictions are billed and asynchronous, so a run waits for approval unless this connection is set to run without asking, and the agent polls for the output.

Find a suitable public model for a task -- image, video, or audio -- and report its latest version id and inputs.
Run a generation and post the result URL back into the thread when the prediction succeeds (billed; after a person approves it)
Check on a long-running prediction when someone asks whether the output is ready yet.

Before you let it write, or keep out: Every Create prediction spends money against the account, and a wrong input can burn a run on garbage. Think before letting it run without asking, and have the agent quote the exact model version and inputs in the thread.

What waits for a person

Reads run on their own. These do not — each one waits in Approvals for an owner or admin unless an admin has set this connection to run writes without asking, and every call toReplicate, approved, rejected or unattended, is logged.

Create prediction POST

What connecting Replicate takes

Replicate uses an API key you create in Replicate and paste in once. A workspace admin sets the connection up; the fields are:

API Token

Credentials are encrypted and bound to your workspace, never read back, and every outbound call passes one guarded fetch — the connector security posture sets out exactly how, and the security page covers the platform around it.

Where Replicate fits

Give an agent Replicate work by writing a message in the inbox, or put its tools in a scheduled flow. What the run cost is on the cost breakdown.

Other ai and media integrations: AnythingLLM, AssemblyAI, D-ID. Or see every application agents can connect to.

See it doing your actual work

Open a demo workspace, connect one app, and give an agent a real job. No card, no sales call.

Try the demo See pricing

We reply to every message within one business day.