Replicate Runtime Skill

Use this skill to generate media via Replicate through the Replicate MCP (mcp__replicate__*).

When To Use

Use Replicate when the user asks to generate an image or a short video from a text prompt, or to run a specific Replicate model. For voice/audio use mcp_elevenlabs; for Chinese-first image/video use mcp_jimeng.

Connection

Authenticates with a Replicate API token. On an auth error, stop and ask the user to fix the token. Generation costs money and time per run — treat each call as billable.

Core Tools

• generate_image — text prompt → image.
• generate_video — text prompt → short video.
• run_model — escape hatch: run any Replicate model by id with arbitrary inputs (use only when the two helpers above don't fit).

Common Recipes

Generate an image

1. Confirm the prompt (and any size/style) with the user. 2. generate_image. 3. Return the result reference/URL.

Run a specific model

1. Identify the model id + its required inputs. 2. run_model with those inputs.

Guardrails

• Each generation is billable and not instant — confirm the prompt before generating; don't speculatively fan out many variations unless the user asked for several.
• Don't silently retry a failed/long generation in a loop — surface status and ask before re-running (re-runs cost again).
• Respect content policy: refuse prompts for disallowed content (e.g. real-person deepfakes, sexual content involving minors, etc.).
• run_model with a wrong input schema wastes a paid run — verify the model's expected inputs first.

Edge Cases & Errors

• Output is async/URL-based — return the produced asset reference; don't fabricate a result if generation hasn't completed.
• Video generation is slower/pricier than image — set expectations and confirm before kicking it off.
• Auth/quota errors → stop and tell the user (token invalid or credits exhausted).