Scenario Agent Skills: Teaching Claude Code, Cursor and Codex to Make Game Art Through MCP

scenario-labs/skills is a collection of Agent Skills that teach AI coding agents to produce game art and media through the Scenario MCP server. Its README says the skills work with Claude Code, Cursor, Codex, Copilot "and 70+ others", covering images, video, audio, textures, skyboxes, 3D assets and custom-trained models. It gained 224 GitHub stars in a day and has about 2,100 in total, which put it on the weekly trending list.

It is a vendor's skill pack: it drives Scenario's own service, so it is both a useful example of how a company ships skills and a product you should evaluate as one. This post covers what is in it, how to install it, and what to check before you let an agent use it.

TL;DR: questions people ask first

What is inside

The README splits the collection in two.

Scenario skills. These run through the Scenario MCP server. The README's skill tables cover a getting-started skill that connects to the server and teaches the "core generation loop" (discover, schema, run, wait, display, download), then groups for finding a direction (references, concept options, moodboards), images (text-to-image, editing, masked inpainting, multi-reference composition, product shots, brand kits, text overlays), and further groups for game assets, sprite animation, textures, skyboxes, 3D, video, video editing and color grading.

Expert tools. The README says these "work differently": they are teams of skills that drive content-creation software and game engines installed on your machine, namely ZBrush, Blender, Maya, Unreal Engine, Unity, Godot and Roblox Studio. They are "distilled from expert tutorials, conference talks, and official documentation", need the application and not the Scenario server, and each family's README records how it was built and how far it was verified. That last note is worth reading before you trust a skill with a real project.

A separate INSTALL file groups skills by role, such as 2D artist, 3D artist and environment artist, and by outcome. For example, the 2D artist bundle is listed at 27 skills and the 3D artist bundle at 15, with example prompts like asking for an 8-frame run-cycle sprite sheet or turning a prop concept into a textured GLB.

A shelf of cards with one pulled out, representing an installable library of agent skills

Example prompts from the install guide

The INSTALL file pairs each role with prompts you can try. These are the project's own examples, not outputs we ran:

• "Give me four directions for a desert outpost, then a moodboard for the one I pick." (the inspiration and image skills)
• "Make an 8-frame run cycle sprite sheet for this knight, and a matching slash effect." (sprite animation and game assets)
• "Put the knight from this sheet in the tavern from this painting, lit like this third image." (multi-reference composition, where each reference has one job)
• "Turn this prop concept into a textured GLB, then show it from the back and from above." (image to 3D and orbit views)
• "Put the dance in this phone clip on my character and export it for Unreal." (video to motion)
• "Make a seamless mossy cobblestone texture set with normal and roughness maps, and a matching overcast skybox." (textures and skyboxes)

The pattern is worth noticing. Each prompt names a deliverable and the inputs that constrain it, and each skill handles the tool-specific detail: model choice, sizing fields, prompt limits, and what to check before declaring the job done. That is the real value of a skill pack. The agent does not need to rediscover the service's quirks on every run, and the checks travel with the instructions.

Why vendor skill packs are spreading

Tool makers have realized that an MCP server alone is not enough. A server exposes functions, but an agent still has to know the order to call them, which settings matter, and how to tell a good result from a bad one. Skills carry that procedural knowledge in plain text. For a company, shipping skills is cheaper than building a custom agent, and it reaches every agent that reads the format, which the README puts at more than 70.

For users, the trade-off is trust. A vendor's skill is written to make its own service work well, which is useful, but it is also written to send traffic to that service. Read the instructions the way you would read a library's code before adding a dependency.

How to install it

The README gives several routes.

# Pick from a menu, grouped by topic (nothing is preselected)
npx skills add scenario-labs/skills

# A single skill
npx skills add scenario-labs/skills --skill scenario

Then connect the server. For Claude Code the README gives:

claude mcp add --transport http scenario https://mcp.scenario.com/mcp

For other MCP clients, add the same URL and sign in with a Scenario account through OAuth, or use an API key. If you are new to the protocol, our guide to what MCP is explains how a client connects to a server and what permissions that grants.

Two practical notes from the README. First, the menu's topic groups cannot be selected from the command line, because the skills installer matches exact skill names only, and the README links an open upstream request for a group flag. For scripts or CI it points you to the default command or a role command from the INSTALL file with -y, which skips every prompt. Second, to stay current run npx skills update where you installed, or re-run the original command.

Pin what you install

The default commands fetch the latest copy of each skill from main. The README offers a safer pattern: pin both the installer version and the commit, taken from a release tag, for example npx skills@1.7.4 add scenario-labs/skills#<commit> --skill <name>, and run it again with a newer commit when you want to move. A skill is text that your agent will follow and may include scripts, so an update to main changes what your agent does. Pinning gives you a state you have read. The project publishes releases through release-please, which builds an open release pull request and, on merge, tags a version and updates the changelog.

We cover the general case in our guide to agent skills. The short version: treat skills like dependencies. Read them, pin them, and update on purpose.

How it compares with other skill packs

Skill packs for creative tools are appearing from the tool makers themselves. Unity published an official plugin, which we covered in Unity ships an official Claude Code plugin with 29 skills, and we wrote a hands-on guide to connecting Claude Code to Unity with Unity MCP. Scenario's pack differs in two ways: most of its skills call a hosted generation service, not a local editor, and the expert-tool families are community-style distillations of tutorials and docs, not first-party plugins from Blender or Unity. That affects how much weight to place on their claims of coverage.

What to check before you use it

1. Account and cost. The Scenario skills need an account. We did not review pricing, so read it first, especially before running video or 3D generation in a loop.
2. What the MCP server can do. An MCP server gives your agent tools. Check what permissions each tool has and keep your agent's file and network access tight.
3. Where your assets go. Prompts, reference images and uploaded product photos go to Scenario's service. If you work with unreleased game art or client material, read the data terms.
4. Output rights. Generated images, video and 3D assets have licensing terms set by the service and the underlying models. Check them for commercial game releases.
5. Skill provenance. The expert-tool families are built from public tutorials and documents. Check each family's README for how far it was verified.
6. Agent safety. Agents that can run local tools in Blender or a game engine can also delete files or run scripts. Use a project copy, version control and a sandbox.

A sensible first test

Install only the core scenario skill, connect the server, and ask for one simple image with a clear reference. Look at how the agent follows the generate-wait-download loop. Then add one more skill, for example image editing, and compare results against doing the task by hand in the Scenario interface. If the agent's results match what you would get yourself with less effort, add more. If they do not, you have learned that cheaply.

What we did not test

We read the README, the INSTALL file and the repository page. We did not install the skills, connect the server, or generate any media, and we did not check Scenario's pricing, terms or the repository's license type. Star counts are as shown on November 4, 2026.

Related reading

• What are agent skills? A complete guide
• What is MCP? The Model Context Protocol guide
• Unity ships an official Claude Code plugin with 29 skills
• Connect Claude Code to Unity with Unity MCP
• OpenMontage: agentic video production for Claude Code
• MoneyPrinterTurbo: AI short-video generator, terms and data risks

Details are accurate as of November 4, 2026. Check the repository for current skills and releases.