Replicate for AI agents
Replicate runs AI models in the cloud through an API, letting developers generate predictions from image, text, and other models without managing servers. Connect it once through Arc0, and your agent, or Claude, ChatGPT and Cursor, can use it through one MCP endpoint, limited to what each user approved.
What agents do in Replicate.
Run a model prediction
Create a prediction from a public model to generate an image, transcript, or other output on demand.
Check hardware before running
List available hardware types before starting a prediction that needs a specific GPU.
Confirm before canceling a job
Get a prediction's status first, then cancel it only after confirming it's the wrong run.
31 Replicate actions, graded by risk.
Every Replicate action is tagged read, write or destructive, so one policy covers the whole app and new actions inherit the right default.
read
2Look things up. Allowed by default.
- replicate.get_predictionGet Prediction
- replicate.searchSearch Models and Collections
write
25Create and change things. Allow, or ask the user first.
- replicate.create_modelCreate Model
- replicate.update_modelsUpdate Model Metadata
- replicate.create_predictionCreate Prediction
- replicate.files_getGet File Details
- replicate.files_listList Files
- replicate.models_getGet Model Details
- replicate.account_getGet Account Information
- replicate.models_listList Public Models
- replicate.files_createCreate File
- replicate.hardware_listList Available Hardware
- replicate.trainings_listList Training Jobs
- replicate.collections_getGet model collection
- replicate.deployments_getGet Deployment Details
- replicate.collections_listList model collections
- replicate.deployments_listList deployments
- replicate.predictions_listList All Predictions
destructive
4Delete, cancel or archive. Ask first, or deny outright.
- replicate.cancel_predictionCancel Prediction
- replicate.files_deleteDelete File
- replicate.trainings_cancelCancel Training
- replicate.deployments_deleteDelete Deployment
Replicate in three steps.
- 01Your users connect ReplicateThey add their Replicate api key on Arc0 Connect, under your brand. It goes straight into the vault.
- 02You set the rulesReads run, writes like “create Model” can wait for the user, and “cancel Prediction” can be denied outright.
- 03Any agent can actYour agent calls Replicate through the Arc0 SDK or MCP, and so can Claude, ChatGPT and Cursor. Every call lands on the audit log.
await arc0.policies.set('replicate', { read: 'allow', write: 'ask', // create_model destructive: 'deny', // cancel_prediction }) # Claude Code: the same connection, one URL $ claude mcp add --transport http arc0 \ https://mcp.arc0.ai/u/u_8f2
How Replicate connects.
Users add their Replicate api key on Arc0 Connect. It is encrypted in the vault, never shown to the model, and each user can rotate or revoke it at any time.
The same Replicate connection serves your agent over MCP and your own backend over REST and the proxy, so a user connects once. How Arc0 handles credentials →
- AUTH
- API key
- CREDENTIALS
- Per-tenant encrypted vault
- MODEL SEES
- Results only, never credentials
- AUDIT LOG
- Every call, on every plan
Use Replicate from any agent.
Replicate and Arc0, answered.
Can I use Replicate with Claude, ChatGPT or Cursor?
Yes. Connect Replicate to Arc0 once, then add your Arc0 MCP URL to Claude, ChatGPT, Cursor, Claude Code or any other remote-MCP client. Each assistant only gets the Replicate actions you allow.
How do users connect Replicate?
Users add their Replicate api key on Arc0 Connect. It is encrypted in the vault, never shown to the model, and each user can rotate or revoke it at any time.
Which Replicate actions can my agent take?
31 in total: 2 read, 25 write and 4 destructive, such as “create Model”. Your policies decide which of them each agent may call.
Can I stop my agent from deleting things in Replicate?
Yes. Actions like “cancel Prediction” are graded destructive. Set destructive actions to deny, or to ask so the user approves each one, and blocked calls still show up on the audit log.
Can my own backend call Replicate too?
Yes. The same Replicate connection is available over REST and through the Arc0 proxy, so your product and your agent share one connection per user.
Plug Replicate into your agent.
Your users connect Replicate once, under your brand. Your agent gets 31 actions behind your policies, with every call on the record.
Free to build · MCP + REST · Audit log on every plan