MCP

A read-only MCP server for game art and scenes

PX-Art runs an MCP server called perfectpixel-scene that lets Cursor, Claude Code, Codex and other MCP-capable clients read your game projects, characters, asset families and scene layouts over JSON-RPC. It exposes twelve tools and no write path: agents can discover, preview and take delivery of assets, but they cannot generate, edit or delete anything.

What MCP is, in one paragraph

The Model Context Protocol is a JSON-RPC convention for connecting an AI client to an external system. The server advertises tools (callable functions with JSON schemas) and resources (addressable documents identified by URI). The client calls initialize, then tools/list or resources/list, then tools/call. That is the whole surface. PX-Art implements the resource and tool capabilities only — no prompts, no subscriptions, no writes.

Why a game-asset MCP is a different problem

Most MCP servers hand back text. Game art is binary, hierarchical and expensive to look at. Three constraints shaped this server:

The twelve tools

Resource URIs

Finished artefacts are read as resources rather than tool calls:

Calling it

The recommended path is list_games → list_game_structure or list_characters → list_asset_families → search_assets → preview_assets → get_asset_downloads. For a whole scene, skip the per-asset steps and call get_scene_bundle.

A tools/call request

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "tools/call",
  "params": {
    "name": "search_assets",
    "arguments": {
      "notebookId": 12,
      "query": "slime",
      "kind": "character",
      "limit": 5
    }
  }
}

The result

Tool results arrive as MCP content items. The text item holds JSON; the shape below is abridged to the fields an agent usually acts on.

{
  "assets": [
    {
      "assetId": 184,
      "notebookId": 12,
      "pageId": 31,
      "uri": "asset://184",
      "name": "slime-walk-south",
      "kind": "action",
      "file": "slime-walk-south.png",
      "frames": 6,
      "w": 768,
      "h": 128,
      "sampling": "nearest"
    }
  ],
  "truncated": false
}

The sampling hint matters: set your engine's texture filter from it, or pixel art arrives blurred. Taking delivery then looks like this:

{
  "assets": [
    {
      "assetId": 184,
      "file": "slime-walk-south.png",
      "mimeType": "image/png",
      "w": 768, "h": 128, "frames": 6,
      "downloadUrl": "https://img.px-art.com/...",
      "expiresAt": 1753240000
    }
  ]
}

Permissions, scoping and audit

1. Read-only by constructionThere is no write, generate or delete tool in the server. This is not a permission setting an agent can escalate past — the capability does not exist.
2. Tokens are per userA token is issued to one account and scoped to that account's own games. It cannot reach another user's projects.
3. Paid status is checked per callEvery call re-validates entitlement server-side rather than trusting a flag captured at token-issue time.
4. Calls are auditedMCP reads are written to an audit trail, including each issue of asset download links.
5. Batch caps are hardEight assets per preview call, fifty per download call. Oversized requests are rejected with a message telling the agent to narrow or split, which it can act on without human help.
6. Editor state stays outScreenshots of the editor and live editor state are not exposed. Agents read data, not your screen.

Read-only MCP is unlocked permanently by any subscription payment or top-up credit purchase. See pricing for what credits cost and what generation spends; formal checkout is still under review and not open yet.

Limits and things to check

• Nothing here generates art. If your agent needs a sprite that does not exist yet, a human generates it in the workspace first — see the sprite sheet generator and the tileset generator.
• Assets are probabilistic output. An agent reading correct metadata about a mediocre sprite still gets a mediocre sprite. Review before shipping.
• Truncation is real. Honour the truncated flag; raise the limit or tighten filters instead of assuming you saw everything.
• Names are not looks. Matching name, role and size do not mean the asset reads correctly in context. When in doubt, call preview_assets.
• Rights still apply. Commercial use is allowed for legally generated assets, but third-party rights, local law and AI-provider restrictions remain your responsibility, and AI output is neither guaranteed exclusive nor guaranteed registrable.
• Prompts are not exposed. The final assembled prompt is internal and is never returned through MCP or any other client-facing surface.

FAQ

Related