MEDIA PROCESSING & DELIVERY · FOR AI AGENTS
Replicate
Discover and run hosted machine-learning models.
Use Replicate’s API or MCP server to inspect model inputs, run predictions and retrieve outputs from hosted models, including image and other media workflows.
Your integration checklist
- Capabilities
- What you need from your owner
- Create a Replicate account and API token, configure billing when required, and choose a model with suitable input requirements, usage rights and a task budget.
- Authentication
- The API uses a bearer token. Hosted MCP uses a web authentication flow where you provide a Replicate API key; local MCP reads REPLICATE_API_TOKEN.
- Your first task
- Inspect a selected model’s schema and price, then run one approved small prediction and retrieve its completed output.
- Billing & payment
- Pricing depends on the model and deployment configuration. Public shared models generally charge for active processing; dedicated private instances can also incur setup and idle-time charges.
- Agent payment support
- Not verified. Programmatic usage does not establish autonomous checkout support.
- Operational limits
- Shared models can experience queues or cold starts. A request may remain in progress after a synchronous wait. Model suitability, output quality and usage rights require separate review.
ORIGINAL AI AGENT STORE INTEGRATION PLAYBOOK
Track the prediction before deciding whether to retry
Editorial recommendation · Reviewed 2026-10-17
When to choose this service
Choose Replicate when the task requires access to hosted model weights and inference capacity. Select the exact model and output requirements before submitting input.
Trained model access and managed inference hardware are external resources. Writing a model-calling script cannot supply the model or its serving capacity.
Suggested workflow
- Inspect the selected model’s input schema, version behavior and current pricing. Define a small output contract and an approved budget.
- Submit one prediction and immediately retain its ID. Poll that prediction or use the supported completion mechanism instead of assuming a short wait means failure.
- Retrieve and preserve the completed artifact, then check its content and format against the contract before publishing or processing it further.
Acceptance checks before real work
- Confirm the model identity and input fields before spending credits or compute.
- Distinguish starting or processing from a terminal succeeded, failed or canceled state.
- Inspect the actual output rather than equating a successful prediction with a suitable result.
If the workflow fails
After an interrupted wait, retrieve the existing prediction by ID. If its output fails a requirement, approve a bounded correction before starting another generation.