Replicate MCP Server – AI Image and Video Models
Replicate's official MCP server gives AI agents access to Replicate's HTTP API for discovering, comparing, and running hosted AI models. Use it for image generation, image editing, upscaling, video generation, model experimentation, and multi-step creative workflows.
Overview
Replicate's official MCP server connects AI agents to Replicate's HTTP API, giving them a structured way to discover, compare, and run thousands of hosted AI models. Although Replicate supports many model types, it is especially useful for design workflows that need image generation, image editing, upscaling, video generation, and creative model experimentation from the same AI client.
What the MCP server enables
Replicate documents MCP access to the operations exposed through its HTTP API. Depending on account access and model availability, an AI agent can:
- Search Replicate models and collections.
- Compare models by metadata, capabilities, and usage details.
- Fetch model information, schemas, versions, and examples.
- Create predictions to run image, video, audio, and other AI models.
- Fetch prediction status and outputs.
- Locate upscalers, image generators, video generators, fine-tuned models, and specialized creative tools.
- Chain model calls, such as generating an image and then upscaling it.
- Use local Code Mode experimentally to execute TypeScript with the Replicate SDK for more complex workflows.
When to use it
Use Replicate MCP when a design or creative-production workflow needs flexible access to many AI models rather than a single image provider. Practical examples include finding the best model for a visual style, generating product or campaign imagery, upscaling assets, producing short video concepts, comparing FLUX or video models, testing new public models, and automating multi-step model workflows from an AI coding tool.
Connection and authentication
Replicate offers a hosted remote MCP server and a local npm package. The hosted option is recommended for most users and is added with the remote server URL:
https://mcp.replicate.com/sse
The hosted flow opens a browser-based authentication page where the user provides a Replicate API token. Replicate stores that token for the server to use on the user's behalf.
For local stdio clients, Replicate publishes replicate-mcp:
npx -y replicate-mcp@latest
The local server requires REPLICATE_API_TOKEN. Keep the real token in a secret manager or environment variable and never place it directly in this YAML.
Key considerations
Running models can incur Replicate usage charges and may consume significant time for large images, videos, or high-end hardware. Model outputs, licenses, safety behavior, pricing, and input schemas vary by model, so agents should inspect model metadata before running production workflows. Hosted remote MCP is easiest and automatically updated. Local execution gives more control and supports experimental Code Mode, but requires Node.js and, for Code Mode, Deno.
Supported Transports
sse
URL: https://mcp.replicate.com/sse
stdio
Command: npx
Args:
-yreplicate-mcp@latest
Frequently Asked Questions
- When should an AI agent use the Replicate MCP server?
- Use it when a workflow needs to discover, compare, or run Replicate-hosted models for creative production, such as image generation, image editing, upscaling, video generation, model testing, or chained model workflows.
- What does the Replicate MCP server add to an AI agent's capabilities?
- It gives the agent structured access to Replicate's HTTP API, allowing it to search models, compare metadata, retrieve schemas, create predictions, and fetch prediction outputs instead of relying only on static model knowledge.
- What can an AI agent access or manage through Replicate MCP?
- Depending on the Replicate account and model availability, the agent can work with model search, model metadata, versions, predictions, prediction results, and supported models for image, video, audio, language, and other inference tasks.
- How is authentication configured for the Replicate MCP server?
- The hosted server uses a browser-based flow where the user provides a Replicate API token. The local stdio server reads the token from REPLICATE_API_TOKEN. Store tokens securely and never commit real credentials to configuration files.
- Which transport should be used for the Replicate MCP server?
- Use the hosted SSE endpoint at https://mcp.replicate.com/sse for the recommended remote setup. Use stdio with the official replicate-mcp npm package when a local server is required or when experimenting with local Code Mode.