Replicate Flux MCP

What awkoy/replicate-flux-mcp MCP server does

The awkoy/replicate-flux-mcp MCP server connects an MCP-compatible client to Replicate for image and model generation. Its curated image tools use black-forest-labs/flux-schnell by default, while SVG requests use recraft-ai/recraft-v3-svg. Both defaults can be changed through environment variables.

The server supports more than one-shot raster generation. Agents can submit a prompt for one image, submit an array of prompts for batch generation, or request several variants from one prompt. SVG generation is available separately for vector-oriented results such as icons, logos, and diagrams. Prediction tools provide access to recent Replicate runs and details for a selected prediction.

How it works

An MCP client starts the server locally through the published npm package and passes the Replicate token through the environment. Tool calls are translated into Replicate API requests. Curated image model overrides are checked against a built-in allowlist, and the SVG tool has its own supported model entry.

For models outside the curated generation workflows, the awkoy/replicate-flux-mcp MCP server provides run_model for allowlisted models and run_replicate_model for a model reference in owner/name[:version] form. The latter can be paired with get_model_schema, which retrieves a Replicate model's OpenAPI input schema and description before an agent constructs the payload. An optional REPLICATE_MODEL_ALLOWLIST can restrict arbitrary model execution.

Generation tools return structured content in addition to human-readable results. Batch and variant calls can send progress notifications when the client supplies a progress token. The repository also describes history resources for image, SVG, and prediction runs, along with prompt templates surfaced by supported clients.

Setup and configuration

Create a Replicate API token and set REPLICATE_API_TOKEN in the MCP client's server configuration. The README shows the package running with npx -y replicate-flux-mcp, including configurations for Cursor, Claude Desktop, and Codex. Smithery and Glama.ai are also documented as hosted setup options.

REPLICATE_IMAGE_MODEL_ID changes the default raster model, and REPLICATE_SVG_MODEL_ID changes the default SVG model. If these variables are omitted, the Flux Schnell and Recraft V3 SVG defaults apply. Model overrides for curated tools still need to satisfy the relevant allowlist. The Codex example also sets a startup timeout, but that is a client configuration rather than a server environment variable.

Tools and capabilities

• generate_image: create one raster image from a prompt.
• generate_multiple_images: process multiple prompts in one request.
• generate_image_variants: create multiple versions of one prompt.
• generate_svg: request vector output from a prompt.
• prediction_list: retrieve recent predictions for the configured Replicate account.
• get_prediction: inspect one prediction by identifier.
• run_model: send raw input to a whitelisted Replicate model, including supported video or restoration use cases.
• run_replicate_model: run a model by its Replicate owner/name reference.
• get_model_schema: inspect the input schema needed by a Replicate model.

The repository identifies the server as MIT-licensed and documents compatibility with Claude Desktop, Cursor, Cline, Zed, and other spec-compliant MCP clients. Running model predictions requires the user's Replicate credentials and may incur Replicate usage costs.