Replicate MCP Server

Replicate hosts thousands of open-source models behind one billing relationship and one API — Flux and SDXL for images, video generators, whisper variants, upscalers, background removers and a long tail of research models. Its MCP server (available hosted at mcp.replicate.com with OAuth) hands your assistant that entire catalogue.

The shape of the workflow: describe what you want, let the assistant pick an appropriate model (or name one), run the prediction, and get the output — then iterate on prompt or parameters in conversation. "Generate a hero image of a mountain lake at dawn, photorealistic" runs Flux; "now make it more cinematic and upscale it" chains a second model onto the first's output. Model chaining through conversation is the quiet superpower: transcribe with one model, summarise with another, illustrate with a third.

Because the assistant can search the catalogue, discovery is included — "what's the best current model for restoring old photos?" returns live candidates with their capabilities, not a stale blog post's opinion. Predictions bill per-run through your Replicate account, typically cents for images; check pricing before video work, which runs hotter.

Against ElevenLabs (audio specialist) and native image tools: Replicate's edge is breadth and openness — when you want a specific open model or the latest research checkpoint, this is where it's runnable without owning a GPU.

Cost and model selection

Replicate bills by the second of hardware time, so which model gets picked matters more than how the prompt is worded.

• Call get_model before the first run of anything unfamiliar. It returns the input schema and hardware tier, which is how you avoid learning a video model's per-run cost from the invoice.
• Predictions are asynchronous: create_prediction returns straight away and get_prediction polls for the result. Rarely-used models cold-start, which can add tens of seconds before any real work happens.
• Pin the version hash when an output is worth reproducing. A bare owner/model reference follows the latest push, and results drift quietly when the maintainer retrains.
• Set a spend limit on your Replicate account before pointing an agent at the catalogue — it is the only hard stop between an enthusiastic iteration loop and a memorable bill.