Quick picks
| If you want… | Do this |
|---|
| Connect Claude Code in one command | claude mcp add --transport http scenario https://mcp.scenario.com/mcp, then sign in with OAuth on first use |
| Connect Cursor | Settings → MCP, add a server with URL https://mcp.scenario.com/mcp (native OAuth/HTTP from Cursor 1.0) |
| Connect Codex CLI | codex mcp add scenario --url https://mcp.scenario.com/mcp, then codex mcp login scenario |
| Teach the agent game-art workflows | Install Scenario's MIT skills: at least scenario + scenario-game-assets (see skills) |
| Batch sprites or icons without guessing models | Ask the agent to recommend → model_schema_get → model_run with dry_run=true first → jobs_wait → asset_download as PNG |
| Ship the art commercially | Use a paid Scenario plan (Starter and up). Free-plan outputs are personal and evaluation use only; see plans |
| Sprites + animation + audio + 3D from one MCP instead | Voidless hosted connector at https://mcp.voidless.dev/mcp (see sprite guide) |
| True pixel-grid art from an MCP instead | Retro Diffusion at https://mcp.retrodiffusion.ai/mcp (see sprite guide) |
What Scenario MCP is
Scenario is a game-first generative art platform (custom style training, 2D, 3D, video and audio). Its MCP server exposes that platform to coding agents over Streamable HTTP at https://mcp.scenario.com/mcp. Scenario's docs say the server covers 99 tools across 750+ models (plus any models you train), with an intelligence layer that recommends models, reads each model's schema and can rewrite prompts. The product page also pitches automated game-asset production: characters, props, environments and UI from the same chat that writes your game code.
Two layers matter for game art:
- The MCP tools (always available once connected): discover teams/projects, search or recommend models, run generations, wait on jobs, display and download assets, check usage.
- Agent Skills (optional, MIT, in scenario-labs/skills): short playbooks that teach the agent which tools to call for sprites, tilesets, textures, skyboxes, 3D and more. Without them the agent can still call tools from schemas; with them it follows Scenario's documented loops.
Install: Claude Code, Cursor and Codex
You need a Scenario account (app.scenario.com) and an MCP-compatible client. OAuth is the recommended auth: no API key passes through the chat. Scenario's connection guide also covers API-key headers for headless/CI setups; treat key wiring as an operator task and never paste secrets into an agent prompt.
Claude Code
claude mcp add --transport http scenario https://mcp.scenario.com/mcp
OAuth starts on first use; run /mcp and pick Scenario to authenticate by hand if needed. Claude apps signed into the same Claude account can also add a custom connector once (claude.ai → Customize → Connectors) and reuse it across Claude Code and Claude Desktop. Prefer the native HTTP transport: Scenario warns that the mcp-remote stdio bridge is the usual source of repeated sign-in prompts.
Cursor
In Settings → MCP (native OAuth and HTTP from Cursor 1.0), add:
{ "mcpServers": { "scenario": { "url": "https://mcp.scenario.com/mcp" } } }
Windsurf (Cascade → Configure MCP) uses the same URL shape. Verify with a free call such as “List my Scenario teams” (teams_list). If a working connection later fails, deactivate and reactivate the server before debugging anything else.
OpenAI Codex
codex mcp add scenario --url https://mcp.scenario.com/mcp
codex mcp login scenario
VS Code: Command Palette → MCP: Add Server → HTTP → the same URL. Other clients (Gemini CLI, Zed, Warp) take the URL through their own MCP config; see Scenario's docs for each.
Pick a team and project first
Scenario organises work into teams (shared workspaces) and projects (assets inside a team). The skills tell the agent to call teams_list, then projects_list, confirm the pair with you, and pass team_id and project_id on every later call. A wrong guess writes into someone else's project. Unattended runs should take the pair from the task or stop and list the choices.
Install the game-art skills
Skills live in Scenario's public GitHub repo under the MIT license. The installer is npx skills add scenario-labs/skills. For game art, Scenario's INSTALL.md groups a ready command (checked November 5, 2026):
npx skills add scenario-labs/skills --skill scenario scenario-game-assets scenario-sprite-animation scenario-side-view-game-kit scenario-blender-2d-animation scenario-textures scenario-skyboxes scenario-3d scenario-image scenario-patina-retexture scenario-walkable-room scenario-video scenario-consistency scenario-model-training scenario-refine-loop scenario-moderation scenario-meshy scenario-rodin scenario-sparc3d scenario-3d-worlds scenario-gemini-image scenario-audio scenario-sonilo
Minimum useful set for 2D game art: scenario (connection + core loop) and scenario-game-assets (sprites, icons, props, tilesets, pixel art, UI, parallax planes, transparent backgrounds, upscaling, PNG export). Add scenario-sprite-animation for walk cycles and VFX sheets, and scenario-textures / scenario-skyboxes when you need those. The repo also has expert-tool families for Blender, Unity, Godot, Unreal and Roblox Studio that drive the local apps rather than Scenario MCP; skip those unless the DCC is installed.
If a skill is missing mid-session, Scenario MCP exposes skill_get so the agent can load it, or you can install it with the same npx skills add command.
The batch loop: recommend → schema → dry_run → generate → download
Scenario's scenario and scenario-game-assets skills document the same core loop. Use it for every batch so the agent doesn't hardcode model IDs (catalogs differ per team) or burn credits on a bad schema:
- Find a model. Call
recommend with the asset need in plain words (sprite, game icon, tileset, isometric, pixel art). Prefer a model whose entry or schema says it outputs alpha when you need transparency, so you skip a billed background-removal pass. Use search only when you already know a model name. Pass include_private_models: true on recommend to include models you trained. - Read the schema. Always call
model_schema_get before model_run. Note the prompt field name, size fields, sample-count parameter and any cost_impact: true fields. Overrunning a cap (for example prompt max_length) is a 400, not a trim. - Price the batch. Run
model_run with dry_run=true first. One dry run prices one leg; multiply across the batch. Scenario's skills warn that a top-level dry_run misplaced into a catalog executor can bill for real — keep dry_run where the direct model_run tool expects it. - Generate. Launch with
wait=false in waves under the concurrency ceiling, relaunch any 429s, then jobs_wait on up to 32 job ids in one call. Never busy-loop job_get. Keep one prompt template and change only the item name across a set (health potion → mana potion), so the style stays locked. - Review and export.
asset_display to inspect; asset_download with format="png" for engine import. Don't paste raw asset URLs into chat. File keepers into collections (Drafts / Sources / Final is Scenario's own Low Thunder example) so the next batch can reuse references.
Worked example from the scenario-game-assets skill: four style-matched potion icons with transparent backgrounds. Prefer a native-alpha model; otherwise generate on a plain background and run a removal pass. Sample-count on one prompt makes variants of one item; different items are separate runs with a fixed wording template.
Plans, credits and commercial rights
From Scenario's pricing page (checked November 5, 2026; annual billing saves 33%):
| Plan | Credits / month | Game-art notes |
|---|
| Free | 50 credits a day, no card | Personal and evaluation use only; no API; no custom model training |
| Starter | 1,500 | Cheapest paid plan with a full commercial license and API access; still no training |
| Pro | 5,000 | Cheapest plan that trains custom style/character models; priority queue + unlimited relaxed generations in the web app when credits run out |
| Max | 10,000 | Teams up to 25, more training capacity |
Shipping rule: regenerate or re-export every asset you ship on a paid plan, and log the tool, plan and date. Free-plan Scenario outputs are not licensed for commercial games. Who owns free vs paid art next to Leonardo is covered in Scenario vs Leonardo.Ai; Steam disclosure wording and a free CSV log are in Steam AI disclosure and steam-ai-asset-log.csv.
When to use Scenario MCP vs PixelLab (or the web app)
- Scenario MCP: you already live in Claude Code / Cursor / Codex, you need style-matched batches across many asset types, or you want the same agent that writes game code to also pull art. Custom-trained models (Pro+) shine here.
- PixelLab: true pixel-grid characters, 4/8-direction rotations and skeleton animation. PixelLab also has its own MCP; use it when the output must be on a real pixel grid. Comparison: PixelLab vs Scenario.
- Scenario web app: still fine for one-off explorations, Icon Set Forge clicks and visual review. MCP wins when the batch is large enough that switching tabs costs more than the prompt.
What we haven't tested
We read Scenario's MCP docs, features page, pricing page and the public skills repo (including scenario and scenario-game-assets) on November 5, 2026. We have not connected a Scenario account from this box, run teams_list, spent credits, or timed a real batch, so we can't vouch for latency, concurrency limits in practice, or output quality from any specific model. Treat the install commands and tool names as documentation facts; replace the “not tested” note once you've run a dry-run batch on your own account.
Sources
Checked November 5, 2026.
- Scenario MCP: Getting Started / docs (99 tools, 750+ models, OAuth, endpoint), features/mcp, pricing
- Skills (MIT): scenario-labs/skills,
skills/scenario/SKILL.md, skills/scenario/references/setup.md, skills/scenario-game-assets/SKILL.md, INSTALL.md (Game art and 3D group) - Case study (vendor narrative, not a how-to we reproduced): Low Thunder — four-day game with Scenario MCP
Related: Scenario vs Leonardo.Ai · how to make game assets with AI · AI game icon and UI generators · best AI sprite generators · AI tools for game art · PixelLab vs Scenario.
Get the Starter Stack ($12) Get the free weekly digest