Skip to main content
Back

Scenario Labs Opens MIT Agent Skills Library Covering 70+ AI Creative Tools

Scenario Labs open-sourced 70+ Agent Skills that teach Claude, Cursor, and Codex to generate production-ready images, video, audio, and 3D through MCP.

Repo, Full Access
  • Scenario Labs open-sourced 70+ Agent Skills for images, video, audio, 3D over MCP
  • Works with Claude Code, Cursor, Codex, Copilot and 70+ other agents via the open Agent Skills format
  • Install with npx skills add scenario-labs/skills --skill "*" and connect mcp.scenario.com
  • Core loop: discover, schema, run, wait, display, download, with dry_run pricing before billed runs
  • Covers Veo, Seedance, Kling, Runway, ElevenLabs, Meshy, Rodin plus Blender, Maya, Unreal, Unity drivers
  • MIT licensed, 1,400+ stars, no version pinning so installs always pull latest main

Scenario Labs has released an MIT-licensed library of Agent Skills for creating images, video, audio, textures, skyboxes, and 3D assets from coding agents. The repository supports Claude Code, Cursor, Codex, Copilot, and, according to its README, more than 70 other compatible agents. The repository has 11 stars.

Scenario provides generative media models through a single platform. Its skills library gives agents the procedural knowledge required to select a model, structure a request, estimate its cost, preserve visual identity, and recover from failed or blocked generations. Those instructions address common problems in automated creative pipelines, where model capabilities, request schemas, reference inputs, and pricing vary widely.

What the agent learns

Each skill centers on a SKILL.md file written in the open Agent Skills format. A compatible agent loads the relevant instructions when a task calls for them, allowing the same workflow to operate across multiple clients without a client-specific rewrite.

The files describe how to use the Scenario MCP server. MCP, short for Model Context Protocol, gives an agent a standard interface for discovering and calling external tools. Scenario performs the generation through that server, while the skill tells the agent which tools to call, which parameters to supply, and how to handle the result.

Install it in two steps

Installation adds the skills to the agent and registers Scenario as an MCP server. The repository documents the following commands for installing every skill and connecting Claude Code:

npx skills add scenario-labs/skills --skill "*"
claude mcp add --transport http scenario https://mcp.scenario.com/mcp

The first command installs the full library through the skills CLI. Developers can also select individual entries, including the core scenario skill. The second command is specific to Claude Code; Cursor, Codex, and other clients require the same endpoint to be registered through their own MCP configuration.

Authentication uses either Scenario OAuth or an API key. The skills carry no separate usage fee, but generation requests are billed to the connected Scenario account. Account credentials and service access therefore remain prerequisites even though the instruction files are open source.

A task-focused catalog

The library groups skills by creative job, allowing an agent to begin with the requested output and then choose an appropriate model or provider.

Category Supported work
Images Text-to-image generation, masked inpainting, product shots, brand kits, storyboards, and text overlays
Game art and 3D Sprites, tilesets, PBR textures, 360-degree skyboxes, and image-to-3D or text-to-3D assets with GLB and FBX export
Video and audio Text-to-video, lip sync, music videos, talking-head clips, scoring, and captioning
Identity and training Character or product continuity, reusable identity libraries, and custom LoRA-style model training
Review Automated quality checks with pass, warn, or fail results and revision suggestions
Provider guidance Dedicated workflows for Veo, Seedance, Kling, Runway, Wan, MiniMax Hailuo, ElevenLabs, Meshy, Rodin, and other model families
Creative tools Workflows involving Blender, Maya, ZBrush, Unreal Engine, and Unity

LoRA-style training creates a compact customization layer for a recurring subject or visual style. That option can help teams reuse an approved character, product, or art direction across later requests without rebuilding the reference set each time.

From prompt to saved asset

The repository’s recommended generation loop gives agents a repeatable sequence for tool selection, pricing, execution, and review:

  1. Discover: Find the relevant Scenario tool and model for the requested medium and output format.
  2. Inspect: Read the tool schema to identify required prompts, references, dimensions, and model-specific settings.
  3. Estimate: Run the request with dry_run enabled to obtain pricing before submitting a billed job.
  4. Generate: Submit the approved request with the selected model and references.
  5. Wait: Poll asynchronous jobs until they complete or return an error.
  6. Deliver: Display the result and download the asset in the required format.
  7. Review: Apply a quality gate and feed concrete revision instructions into another run when needed.

Model-routing guidance reduces the research required to choose among image, video, audio, and 3D providers. The skills also document recovery paths for blocked generations, helping the agent adjust a request instead of repeatedly submitting the same failing prompt.

Guardrails for cost and continuity

  • Preflight pricing: Dry runs expose the expected charge before the agent starts a billed generation.
  • Baseline-plus-delta prompts: A stable description preserves the subject’s defining traits, while a smaller delta specifies changes such as pose, camera angle, clothing, or setting.
  • Identity libraries: Reusable reference collections give later generations a consistent source for characters and products.
  • Quality gates: Pass, warn, and fail verdicts turn review criteria into revision instructions for the next request.

Where the workflows fit

Teams producing repeated variations gain the clearest benefit: game studios generating matched asset sets, marketing groups adapting an approved campaign, and independent developers creating props or textures without leaving their coding agent. The repository gives examples such as:

  • Generate four style-matched potion icons with transparent backgrounds for my RPG inventory.
  • Turn this concept sketch into a 3D prop and let me preview it before exporting the GLB.

These requests combine several operations that agents often mishandle when working from API documentation alone, including model selection, reference handling, format selection, job polling, and file delivery.

Dependencies can complicate installs

Some pipeline skills call sibling skills, so a narrowly selected installation can leave required references unavailable. The repository addresses this with longer installation commands that bundle director skills with their dependencies. Developers choosing individual skills should inspect those relationships before relying on a workflow.

The documented installation flow pulls the latest files from main and does not provide a pinned release in the command. A skill update can therefore change agent behavior between installations. Teams with reproducibility requirements can record the reviewed commit or vendor a known copy of the relevant files.

Final output quality still depends on the selected model, source references, prompt quality, service availability, and the host agent’s MCP implementation. External creative tools may also require their own installations, licenses, and project configuration.

Portable procedures for changing APIs

A shared SKILL.md can carry the same creative procedure across compatible agents, reducing dependence on one coding client. Procedural files also suit fast-changing media APIs because maintainers can update model choices, schemas, and recovery guidance without retraining an agent.

Teams evaluating the library can start with a fixed benchmark task, run the pricing preflight, inspect the generated request, and compare consistency across several outputs. That test will show whether the repository’s routing and review instructions improve the specific pipeline before broader adoption.