$ loading_
通过 Composio MCP 自动化执行 Replicate 模型调用、文件上传与预测管理。
复制安装指令，让 AI 自动完成配置 · 推荐新手
请帮我安装 askskill 上的 "Replicate Automation" 技能： 1. 下载 https://raw.githubusercontent.com/ComposioHQ/awesome-claude-skills/master/composio-skills/replicate-automation/SKILL.md 2. 保存为 ~/.claude/skills/replicate-automation/SKILL.md 3. 装好后重载技能，告诉我可以用了
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://rube.app/mcpUse 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.
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.
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
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)
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)
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
| 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. |
…
通过 Rube MCP 自动执行 Jobnimbus 任务与流程操作，提升业务处理效率。
通过 Rube MCP 自动化执行 AI/ML API 任务，并先检索最新工具 schema。