How to Use AI Agents to Manage Your Game Asset Catalog

Your assistant can use your tools

An AI assistant used to be limited to what it had read during training. The Model Context Protocol (MCP) changed that: an assistant like Claude can now connect to an external tool and perform real operations against it, reading data, creating records and running searches under the same permissions as a human user.

PolyDrobe ships an MCP server that exposes the catalog operations. This post covers connecting an assistant to your project and the workflows you can run once it is wired up.

Setting up the connection

What you need

• A PolyDrobe account with at least one project
• An assistant that speaks MCP (Claude Desktop, Claude Code, or any MCP-compatible client)
• A Bearer token from PolyDrobe's OAuth endpoint at /o/token/

Connecting Claude

Add PolyDrobe to your Claude configuration as an MCP server:

{
  "mcpServers": {
    "polydrobe": {
      "url": "https://polydrobe.app/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN_HERE"
      }
    }
  }
}

Claude discovers the available tools on connection and starts using them in conversation. You never list the tools yourself: Claude reads their descriptions off the server and picks the ones your request needs.

Workflows you can run today

Each example below shows what you say and what the assistant does with it.

Browse your catalog

You say:

"List all my projects and show me what is in the Vanguard Skins project."

What happens: The assistant calls list_projects, finds Vanguard Skins by name, then calls get_project for the full details: categories, rarities, statuses and summary counts. You get the overview without opening a browser tab.

Search for specific assets

You say:

"Find all helmets with Epic rarity in the Vanguard Skins project."

What happens: The assistant calls list_assets with category and rarity filters, or hands the whole thing to search. Back come the matching assets with variant counts, statuses and metadata.

That answer lands in seconds during a planning meeting, where the alternative is opening the project, setting two filters and counting rows while everyone waits.

Create assets and variants

You say:

"Create a new asset called Phantom Jacket in the Outerwear category with Rare rarity. Add two variants: Phantom Jacket Obsidian at 800 coins, and Phantom Jacket Chrome at 1200 coins with Epic rarity."

What happens: The assistant calls create_asset with the name, category and rarity, then create_variant twice with the parent asset, names, prices and rarity values. Three API calls out of one sentence.

Brainstorming is where this pays off. You describe the asset while the idea is fresh, and it lands in the catalog without a detour through four form screens.

Check release readiness

You say:

"Give me a release readiness report for Release 2.1. How many variants are Approved, how many are still In Progress, and which ones are blocked in Review?"

What happens: The assistant filters list_assets by release version, walks the variants, groups them by status and reports:

• 47 variants Approved
• 12 variants In Progress
• 3 variants in Review, with names and parent assets

Ask the question in a release standup and the answer reflects the catalog as of that minute.

Compare variants

You say:

"Show me all Legendary-rarity variants across the project with their prices. Are any priced below 2000 coins?"

What happens: The assistant queries the catalog, filters by rarity, pulls the price off each variant and flags the ones under the threshold. The same pricing audit by hand means an export and a spreadsheet formula.

Generate documentation

You say:

"Write a summary of all Season 3 Battle Pass items including asset names, variant counts, rarity distribution, and total item count."

What happens: The assistant searches for items tagged Season 3 or Battle Pass, aggregates them, and formats a summary you can paste into a design doc, a Slack thread or a publisher email.

Bulk edits across many variants

You say:

"Create ten color variants of the Phantom Jacket (Sapphire, Ruby, Emerald, Obsidian, Ivory, Bronze, Copper, Jade, Violet, Amber) all priced at 800 coins with Rare rarity."

What happens: The assistant calls bulk_create_variants once with the full list, and ten variants land in one round trip. bulk_update_variants and bulk_delete_variants cover mass status changes, repricing and cleanup.

Manage tags, categories and releases

You say:

"Add a new category called Cyber Ops under Outerwear, create a Legendary rarity with the color #ffb347, and tag every Phantom Jacket variant with the Season 4 tag."

What happens: The assistant calls create_category, then create_rarity, then add_variant_tags on the matching variants. The classification tools cover categories, tags, rarities, releases, currencies and asset statuses, which is the same vocabulary you maintain in project settings.

Export and import catalog data

You say:

"Export the Vanguard Skins project as JSON so I can review it offline, then re-import the edited file when I am done."

What happens: export_project returns a JSON snapshot and import_project_data applies your edited payload back. Migrations, offline editing and moving a curated subset of assets between projects all run through this pair.

Where the assistant stops

MCP enforces the permission model of the web interface. A Viewer's token reads and searches. An Editor's token creates and updates assets, variants, classifications and tags. An Owner's token adds team member management, project deletion and the other destructive operations.

The MCP server does not cover:

• Uploading full-resolution 3D model files (set_variant_thumbnail sets a thumbnail; textures and meshes go through the web interface)
• Authentication and account management (sign-up, password reset, billing actions)
• Stripe and workspace billing changes

The rest of the web interface is exposed as tools: full CRUD on projects, assets and variants, classification management, tag assignment, member roles, activity logs and bulk operations.

Building automation on top of MCP

The server serves scripts as readily as it serves a chat window:

Pre-release validation. A CI step connects to the MCP server, queries the variants targeted at the next release, and fails the build while any of them sit in In Progress.

Slack integration. A bot answers "What shipped in Release 2.0?" by querying the catalog and posting the summary to the channel.

Content planning. A script pulls the current catalog, compares it against a target such as three Legendary variants per category, and writes a gap report.

Cross-tool sync. An agent reads asset data from PolyDrobe and opens the matching Jira tickets for production work, so the catalog and the task tracker agree.

Getting started

1. Log in to PolyDrobe with at least one project holding some assets.
2. Move to a plan that includes MCP access. The free plan does not carry it.
3. Generate a Bearer token from the OAuth endpoint.
4. Add the MCP server to your assistant's configuration.
5. Open with "List my projects" and build from there.

The pricing page shows which plan includes MCP access alongside the rest of the feature matrix.

Key takeaways

• PolyDrobe's MCP server lets an AI assistant browse, create, update and delete assets, variants and classifications in plain language.
• The workflows that pay off first are catalog queries, batch creation, release readiness checks, pricing audits, bulk variant edits and data export and import.
• Authentication runs on OAuth 2.1 under the same role model as the web interface, so Viewer, Editor and Owner tokens behave exactly as they do in the UI.
• Scripts use the same server: CI validation, Slack bots, content planning and cross-tool sync.
• MCP access comes with the Pro plan, and any MCP-compatible assistant can connect to it.