runway-dev-models

Build, modify, debug, or verify Runway model generation in an application: discover accessible models and constraints with MCP, implement SDK calls, and wire inputs and outputs into the product UI. Use with +runway-dev. Not for Model Routers, Characters, recipes, or agent-side generate scripts.

源文件：runwayml/skills/skills/runway-dev-models/SKILL.md

Runway Dev — Models

Companion: Use +runway-dev for shared guidance when available. If it is not installed, inspect the workspace, probe RUNWAYML_API_SECRET without printing it, and read current docs. Encourage connecting Dev MCP for live model access and constraints. If the user declines or cannot connect, continue from current docs and existing model configuration.

Goal

Help the user choose and integrate a model endpoint (/v1/text_to_video, /v1/text_to_image, etc.) across their application, including its backend call and existing UI. Verify changes with one working SDK call when safe.

MCP tools

list_models — discover models the selected project can call and read each model's inputConstraints.
get_credit_balance — check budget before billable verification.
get_task — inspect or debug an existing task, not replace SDK wait helpers in application code.

Workflow

Inspect the existing application. Confirm the user experience, target modality, and where generation inputs and outputs belong.
If the task requires model selection, access verification, or current constraints, call list_models with { projectId, endpoint }. If existing code pins a model, proceed from current docs unless live access must be verified.
Follow the endpoint docs linked by llms.txt; do not infer request fields from another model.
Keep RUNWAYML_API_SECRET behind the application's server boundary.
Implement or update the SDK call by chaining .waitForTaskOutput() in Node or .wait_for_task_output() in Python directly from the create call.
If the application has a UI, wire its controls to the backend and render loading, error, and generated-output states.
When verification is appropriate, submit one test generation. Present the result and offer to persist output before its signed URL expires.

Input media

Follow the current input docs linked by llms.txt: use a public HTTPS URL, a small data URI, or an ephemeral upload.
Send browser-selected files to the application's server, then use the SDK upload helper and pass its runway:// URI to generation. Local filesystem paths cannot be API inputs.
Do not accept arbitrary remote URLs from clients. Prefer uploads or allowlisted origins.
Ephemeral inputs and generated output URLs expire; persist anything the application must retain.

Do not

Guess model ids, ratios, or durations from memory or other models.
Put API keys in frontend bundles.
Add manual polling when the SDK wait helper fits, or resubmit on transient read errors.

Docs

Index: https://docs.dev.runwayml.com/llms.txt
Setup: https://docs.dev.runwayml.com/guides/setup/