How to give an AI coding agent a 3D asset library over MCP

Install

The server speaks MCP over Streamable HTTP at a single URL, and browsing needs no key or account. Add it once, then restart the agent's session so it sees the tools.

Claude Code

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

Add --scope user to have it in every project; /mcp inside a session lists the tools.

Cursor

{
  "mcpServers": {
    "gaitkit": { "url": "https://gaitkit.com/mcp" }
  }
}

Codex

codex mcp add gaitkit --url https://gaitkit.com/mcp

VS Code (Copilot agent mode)

{
  "servers": {
    "gaitkit": { "type": "http", "url": "https://gaitkit.com/mcp" }
  }
}

Other clients, including Windsurf, Claude Desktop and claude.ai custom connectors, take the same URL. The agents page has each one's config file.

The five tools

• search_assets (query, class, legs, max_triangles, limit): finds specimens and grades its own answer as ok, poor or none. Each result has its size in metres, triangle count, knob names, motion (walk, drive, fly, swim, still) and tier.
• get_asset (id): the full record: knob schema with a description of every knob, presets, night-light zones, rigged parts, gait data for walkers, file URLs and paste-ready three.js code for its kind of motion.
• get_variant (id, want): plain words in, a createAsset({...}) call out. It understands colours, sizes, counts, features, option and preset names, in English or Turkish, and lists the words it could not place.
• report_need (need, context): logs something the catalogue does not have. The request queue decides what gets built next.
• who_am_i: which account an API key belongs to, its plan and its CDN token.

The tool descriptions tell the agent to search first, then fetch details, and to be honest when coverage is poor. You do not need a special prompt.

What an exchange looks like

A real exchange against the live server, shortened. The tool answers are what the server returns; note that the word it cannot act on comes back as unknown_words instead of being guessed at.

> Add a white fox with big ears that walks around the meadow. Make it look sleepy.

● gaitkit · search_assets({ "query": "fox", "limit": 3 })
  coverage: ok · fox (creature, walk, 4 legs, 2572 tris, free), giant-bat, sandbag-nest

● gaitkit · get_variant({ "id": "fox", "want": "white fox with big ears, sleepy" })
  confidence: high
  call: createAsset({"coat":"arctic","ears":1.8})
  unknown_words: ["sleepy"]

● Write(src/meadow/fox.js)
  import { createAsset } from 'https://gaitkit.com/cdn/fox.mjs';
  import { walker } from 'https://gaitkit.com/runtime/walk.mjs';
  const fox = createAsset({ coat: 'arctic', ears: 1.8 });
  const gait = walker(fox, { ground: raycastGround([meadow]) });

  Done: an arctic fox with big ears, walking on the meadow mesh.
  The fox has no "sleepy" option, so I left that out.

The agent writes code against the same public URLs a person would use: the module from /cdn/{id}.mjs and the walking runtime from /runtime/walk.mjs. The fox is a free specimen, so that code runs for anyone.

When the model is not there

A 3D library is only useful to an agent if it says no when it should. search_assets scores its own results: a hit on a name or tag is ok; matches that only touch descriptions are poor; nothing is none. On poor or none the response carries a note telling the agent to say so and to offer report_need, rather than passing off a robot spider as a crab.

> I need a kangaroo that hops.

● gaitkit · search_assets({ "query": "kangaroo" })
  coverage: poor · note: Nothing fits well. Tell the user honestly and offer report_need so it gets built.

  The catalogue has no kangaroo; the nearest results are unrelated. Want me to request one?

> Yes.

● gaitkit · report_need({ "need": "kangaroo that hops", "context": "outback scene, three.js" })
  recorded: true · Logged. Requests decide what gets built next.

Asking for the same thing twice keeps the first place in the queue rather than adding a duplicate. The same queue collects requests from the website's search box.

Free files, Pro files and keys

• No key needed to browse. Search, metadata, variants and every free specimen's files are open. Each result says tier: free or tier: pro.
• Pro files need a plan. Create an API key on your account and pass it as a header. get_asset then returns Pro file links that carry your CDN token.

claude mcp add --transport http gaitkit https://gaitkit.com/mcp --header "Authorization: Bearer gk_live_..."

Keep the API key (gk_live_…) in the agent's config, never in shipped code. The CDN token (gkc_…) only reads files, so it may go into a game or web app. Without a key the server allows 120 calls a minute per address; an agent that wants the whole catalogue should fetch /catalog.json once.

Without an MCP client

The endpoint is plain JSON-RPC 2.0 over POST /mcp, so any script can call it:

curl -s https://gaitkit.com/mcp -H 'content-type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search_assets","arguments":{"query":"fox","limit":3}}}'

The same operations exist as REST: GET /api/assets?q=fox, GET /api/assets/fox, GET /api/assets/fox/variant?want=white+big+ears and POST /api/demand. A short map of the site for language models lives at /llms.txt.

Questions