ApparelHub MCP server

@apparelhub/mcp-server

Workflow-level MCP tools that let an AI agent run an ApparelHub store end to end: set it up, design apparel, build products, list them across sales channels, and work the orders and listings that come back. Each tool wraps the ApparelHub Agent API and bakes in the platform's hard-won production lessons, so the agent gets correct behavior for free instead of learning the gotchas itself.

There are two ways to run it: a hosted connector you authorize with OAuth (recommended, no API key to handle) and this npm package you run yourself with a key. Both serve the identical tool surface.

Status: early access. The npm package is pre-1.0 while the surface stabilizes. The

agent-facing tool surface is v1 and is the contract we keep stable (see

CHANGELOG.md and Versioning).

What makes this different

A thin wrapper around a REST API just renames HTTP calls. These tools are at the workflow level: one ship_product call resolves variants, generates and waits for a mockup (through the waiting until its images are actually published), creates the product with the right field names, adds every variant, associates it with a store, and syncs to fulfillment and channels in the correct order, refusing a negative-margin price and warning on known variant traps along the way. The scar tissue lives in the code, not in your agent's context.

Connect

Prefer the hosted connector. Authorizing it provisions the API key for you, so there is no credential to create, paste, rotate, or leak into a config file. Reach for the self-run package only when your client cannot speak remote MCP or OAuth, or when you specifically want the server running on your own machine.

Recommended: the hosted connector

https://mcp.apparelhub.ai

Add that URL to any MCP client that supports remote servers, sign in to ApparelHub, and approve. You never handle an API key. Approving the grant provisions a connector key on your account (or reuses the one you already have), and the hosted server resolves your session to it on every call. Nothing to paste, nothing to rotate by hand, nothing sitting in a dotfile.

The hosted server also carries the imaging toolchain the design and quality tools need, so transparency keying, image statistics, and OCR all work with no local dependencies at all: no Node, no Python, no Pillow, no tesseract.

Claude Code

claude mcp add --transport http apparelhub https://mcp.apparelhub.ai

Then run /mcp and choose Authenticate.

Any client that reads an MCP config file

{
  "mcpServers": {
    "apparelhub": {
      "type": "http",
      "url": "https://mcp.apparelhub.ai"
    }
  }
}

claude.ai — add it as a custom connector under Settings → Connectors, paste the same URL, then authorize when prompted.

The bare origin above is the server's canonical address: it is the resource identifier its own OAuth metadata advertises. A path is not routed on, so an existing config pointing at https://mcp.apparelhub.ai/mcp keeps working and needs no change.

What the handshake actually does

Shortened. The full README is on GitHub.