Publish AI-built games straight from your agent

How the flow works

1. 1. Upload

   Your agent calls upload_game. GameFoundry stores the HTML and returns a preview URL. The URL only renders for you (signed in as the owner) - nobody else can see the game yet, and it's not on the public gallery.

2. 2. Preview & iterate

   Open the preview URL in a browser. Play the game. Ask your agent to fix things - it calls update_game with the new HTML, no moderation, no re-publish. Repeat until it works.

3. 3. Submit for review

   When you're happy, ask your agent to submit. It calls submit_game, the game goes into human review, and a GameFoundry moderator approves it. Once approved it appears on /creations for everyone.

4. 4. Later

   Push updates any time - update_game on a live game re-runs moderation. Take it off the gallery with unpublish_game to keep iterating. Remove it entirely with delete_game.

What your agent can do

Seven tools. Everything is scoped to the user who minted the token - an agent never sees another user's creations.

• upload_game - a complete HTML document → new DRAFT creation + owner-only preview URL. Not public. No moderation.
• update_game - replace HTML, title, or description on a draft (skips moderation) or on a published game (re-runs moderation).
• submit_game - run moderation, publish to the public gallery.
• unpublish_game - take a live game off the gallery, revert to draft state.
• list_my_games - paginated list of the caller's own creations (drafts + published).
• get_game_stats - title, status, like count, and share URL for one creation.
• delete_game - soft deletes one of the caller's creations.

Setup - 3 minutes

1. Set a public handle at /dashboard/account. This is what appears on your creations. Skipping this step leaves uploads failing with a friendly error until you set one.
2. Connect your MCP client using one of the paths below.
3. In your agent, say: “Upload this game to GameFoundry as a draft.” Then preview the URL. When you're happy, say “Submit it.”

Easiest: OAuth (Claude Desktop, ChatGPT, claude.ai)

The one-click setup. No tokens to copy, no config file to edit. A browser tab opens, you approve, done.

Command line clients: personal access token

Claude Code, Cursor, VS Code and other CLI / config-driven clients use a bearer token. Mint one at /dashboard/api-tokens (shown once, copy it), then paste it into one of the snippets below where marked gf_mcp_REPLACE_ME.

claude mcp add --transport http gamefoundry https://www.gamefoundry.net/api/mcp \
  --header "Authorization: Bearer gf_mcp_REPLACE_ME"

{
  "mcpServers": {
    "gamefoundry": {
      "type": "http",
      "url": "https://www.gamefoundry.net/api/mcp",
      "headers": {
        "Authorization": "Bearer gf_mcp_REPLACE_ME"
      }
    }
  }
}

{
  "servers": {
    "gamefoundry": {
      "type": "http",
      "url": "https://www.gamefoundry.net/api/mcp",
      "headers": {
        "Authorization": "Bearer gf_mcp_REPLACE_ME"
      }
    }
  }
}

Things worth asking

Rules and limits

• 10 writes per hour per token (uploads, updates). Enough for iterating by hand, low enough to catch a runaway loop.
• Game HTML must fit in ~1.5 MB. Everything must be inline - the embed CSP blocks external network requests except for a pinned three.js CDN.
• Only PUBLISHED creations are visible to anyone but the owner. Drafts are private; a preview URL sent to someone else returns a 404 unless they're signed in as the owner.
• Content is moderated for a 13+ audience before it appears in the public gallery.
• A token acts on behalf of exactly its owner. Revoke it any time from the dashboard - the next agent call fails immediately.
• The endpoint is stateless HTTP over POST. JSON-RPC 2.0. No SDK - any MCP-compatible client that accepts a bearer token works.