home / skills / composiohq / awesome-claude-skills / replicate-automation

replicate-automation skill

safe

/replicate-automation

This skill automates Replicate AI model workflows by running predictions, uploading inputs, inspecting schemas, and tracking history via Composio MCP.

npx playbooks add skill composiohq/awesome-claude-skills --skill replicate-automation

Review the files below or copy the command above to add this skill to your agents.

Files (1)

SKILL.md

5.2 KB

---
name: Replicate Automation
description: "Automate Replicate AI model operations -- run predictions, upload files, inspect model schemas, list versions, and manage prediction history via the Composio MCP integration."
requires:
  mcp:
    - rube
---

# Replicate Automation

Automate your Replicate AI model workflows -- run predictions on any public model (image generation, LLMs, audio, video), upload input files, inspect model schemas and documentation, list model versions, and track prediction history.

**Toolkit docs:** [composio.dev/toolkits/replicate](https://composio.dev/toolkits/replicate)

---

## Setup

1. Add the Composio MCP server to your client: `https://rube.app/mcp`
2. Connect your Replicate account when prompted (API token authentication)
3. Start using the workflows below

---

## Core Workflows

### 1. Get Model Details and Schema

Use `REPLICATE_MODELS_GET` to inspect a model's input/output schema before running predictions.

```
Tool: REPLICATE_MODELS_GET
Inputs:
  - model_owner: string (required) -- e.g., "meta", "black-forest-labs", "stability-ai"
  - model_name: string (required) -- e.g., "meta-llama-3-8b-instruct", "flux-1.1-pro"
```

**Important:** Each model has unique input keys and types. Always check the `openapi_schema` from this response before constructing prediction inputs.

### 2. Run a Prediction

Use `REPLICATE_MODELS_PREDICTIONS_CREATE` to run inference on any model with optional synchronous waiting and webhooks.

```
Tool: REPLICATE_MODELS_PREDICTIONS_CREATE
Inputs:
  - model_owner: string (required) -- e.g., "meta", "black-forest-labs"
  - model_name: string (required) -- e.g., "flux-1.1-pro", "sdxl"
  - input: object (required) -- model-specific inputs, e.g., { "prompt": "A sunset over mountains" }
  - wait_for: integer (1-60 seconds, optional) -- synchronous wait for completion
  - cancel_after: string (optional) -- max execution time, e.g., "300s", "5m"
  - webhook: string (optional) -- HTTPS URL for async completion notifications
  - webhook_events_filter: array (optional) -- ["start", "output", "logs", "completed"]
```

**Sync vs Async:** Use `wait_for` (1-60s) for fast models. For long-running jobs, omit it and use webhooks or poll via `REPLICATE_PREDICTIONS_LIST`.

### 3. Upload Files for Model Input

Use `REPLICATE_CREATE_FILE` to upload images, documents, or other binary inputs that models need.

```
Tool: REPLICATE_CREATE_FILE
Inputs:
  - content: string (required) -- base64-encoded file content
  - filename: string (required) -- e.g., "input.png", "audio.wav" (max 255 bytes UTF-8)
  - content_type: string (default "application/octet-stream") -- MIME type
  - metadata: object (optional) -- custom JSON metadata
```

### 4. Read Model Documentation

Use `REPLICATE_MODELS_README_GET` to access a model's README in Markdown format for detailed usage instructions.

```
Tool: REPLICATE_MODELS_README_GET
Inputs:
  - model_owner: string (required)
  - model_name: string (required)
```

### 5. List Model Versions

Use `REPLICATE_MODELS_VERSIONS_LIST` to see all available versions of a model, sorted newest first.

```
Tool: REPLICATE_MODELS_VERSIONS_LIST
Inputs:
  - model_owner: string (required)
  - model_name: string (required)
```

### 6. Track Prediction History and Files

Use `REPLICATE_PREDICTIONS_LIST` to retrieve prediction history, and `REPLICATE_FILES_GET`/`REPLICATE_FILES_LIST` to manage uploaded files.

```
Tool: REPLICATE_PREDICTIONS_LIST
  - Lists all predictions for the authenticated user with pagination

Tool: REPLICATE_FILES_LIST
  - Lists uploaded files, most recent first

Tool: REPLICATE_FILES_GET
  - Get details of a specific file by ID
```

---

## Known Pitfalls

| Pitfall | Detail |
|---------|--------|
| Model-specific input keys | Each model has unique input keys and types. Using the wrong key causes validation errors. Always call `REPLICATE_MODELS_GET` first to check the `openapi_schema`. |
| File upload encoding | `REPLICATE_CREATE_FILE` requires base64-encoded content. Binary files treated as text (UTF-8) will fail with decode errors. |
| Public vs deployment paths | Public models must be run via `REPLICATE_MODELS_PREDICTIONS_CREATE`. Using deployment-oriented paths causes HTTP 404 failures. |
| Sync wait limits | `wait_for` supports 1-60 seconds only. Long-running jobs need async handling via webhooks or polling `REPLICATE_PREDICTIONS_LIST`. |
| Image model constraints | Image models like flux-1.1-pro have specific constraints (e.g., max width/height 1440px, valid aspect ratios). Check the model schema first. |
| Stale file references | Heavy usage creates many uploads. Routinely check `REPLICATE_FILES_LIST` to avoid using stale `file_id` references. |

---

## Quick Reference

| Tool Slug | Description |
|-----------|-------------|
| `REPLICATE_MODELS_GET` | Get model details, schema, and metadata |
| `REPLICATE_MODELS_PREDICTIONS_CREATE` | Run a prediction on a model |
| `REPLICATE_CREATE_FILE` | Upload a file for model input |
| `REPLICATE_MODELS_README_GET` | Get model README documentation |
| `REPLICATE_MODELS_VERSIONS_LIST` | List all versions of a model |
| `REPLICATE_PREDICTIONS_LIST` | List prediction history with pagination |
| `REPLICATE_FILES_LIST` | List uploaded files |
| `REPLICATE_FILES_GET` | Get file details by ID |

---

*Powered by [Composio](https://composio.dev)*

Overview

This skill automates Replicate AI model operations through the Composio MCP integration. It handles model inspection, file uploads, running predictions (sync or async), listing versions, and managing prediction and file history. Use it to safely build workflows that call image, audio, video, and LLM models hosted on Replicate.

How this skill works

The skill exposes toolkit actions that query model metadata and openapi schemas, upload base64-encoded files, create predictions with optional synchronous wait or webhooks, and list prediction/file history. Before invoking a model, you inspect its schema to construct valid inputs. For long jobs you either omit the short wait window and use webhooks or poll the predictions list.

When to use it

You need to inspect a model's input/output schema before calling it.
You want to run quick predictions synchronously (1-60s) or long jobs asynchronously with webhooks.
You must upload images, audio, or other binary inputs for model inference.
You need to paginate and review prediction history for debugging or auditing.
You want to enumerate model versions or read a model's README for usage notes.

Best practices

Always call REPLICATE_MODELS_GET and check openapi_schema before building inputs.
Base64-encode binary files and provide correct content_type when using REPLICATE_CREATE_FILE.
Use wait_for only for fast models; use webhooks or polling for long-running tasks.
Clean up or monitor uploaded files regularly to avoid stale file_id references.
Respect model-specific constraints (image dimensions, aspect ratios) documented in the schema/readme.

Example use cases

Generate images by inspecting flux-1.1-pro schema, uploading a reference image, then creating a prediction with prompt and image file_id.
Run LLM inference by reading a model README, building the request payload from the model schema, and waiting synchronously for short responses.
Process audio: upload audio.wav via REPLICATE_CREATE_FILE, then call a speech model and receive transcription asynchronously via webhook.
Audit model usage by listing predictions and filtering results or logs for billing and debugging.
Compare versions by listing model versions and running the same input across recent releases to evaluate output differences.

FAQ

What format must files use when uploading?

Files must be base64-encoded strings with a filename and appropriate content_type; binary-as-text will cause decode errors.

When should I use wait_for vs webhooks?

Use wait_for (1-60s) for fast models when you need immediate results. For long-running jobs, omit wait_for and use webhooks or poll the predictions list.