Run your marketplace from Claude

mongrl is an MCP server. Connect an AI assistant to your marketplace and it works in the console as you: it finds listings waiting for review, reads an order’s history, drafts support replies and summarises a month of sales, under the same rules as your console account.

	Address
Live	https://api.mongrl.com/mcp
Test	https://api.staging.mongrl.com/mcp

There’s nothing to install and no key to copy. The first time an assistant connects, it sends you to the console to sign in and approve it.

Connect an assistant

Claude (claude.ai, the desktop app): Settings › Connectors › Add custom connector, and paste the address.

Claude Code:

claude mcp add --transport http mongrl https://api.mongrl.com/mcp

Then run /mcp in Claude Code and choose mongrl to sign in.

ChatGPT: add a connector (in developer mode) with the address as its MCP server URL.

Cursor, in .cursor/mcp.json:

{ "mcpServers": { "mongrl": { "url": "https://api.mongrl.com/mcp" } } }

VS Code, in .vscode/mcp.json:

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

Any other client that supports remote MCP servers with OAuth works the same way.

What you approve

When an assistant connects, the console shows what it asks for:

Permission	What the assistant can do
console:read	See listings, orders, users, messages, pages, vouchers, settings and reports. Everyone on the team can grant it.
console:write	Make changes: approve and edit listings, move orders on, reply to support, edit pages, create vouchers. Only admins can grant it.

• The assistant acts as you, in one marketplace (the one you pick when you approve). It can never do more than your role allows, and if your role changes or you leave the team, so does what it can do.
• An assistant granted only console:read isn’t offered the tools that change things.
• Assistants never reach API keys, sign-in, the team or other connected apps.
• Settings › Connected apps lists what you’ve connected. Disconnecting cuts the assistant off within 30 seconds.

What it can do

Tools, in groups:

• Listings: search, read, create for a seller, edit, approve, send back, turn down, put on and take off sale, offer a price, set price drops, commission and logistics, and draft a description with AI.
• Orders: list, read, move on (mark delivered, cancel, resolve a dispute) and book delivery times.
• Users: search, read, make accounts, block and unblock.
• Conversations: list, read and reply to support.
• Pages: list, read, start, save, publish, unpublish, restore versions and draft text with AI.
• Vouchers, categories, catalogue terms, locations, newsletter sign-ups, settings and the sales report.

Each tool tells the assistant whether it only reads, whether it’s safe to repeat, and whether it’s destructive: whether it emails someone, changes the shop or is hard to undo. Good clients ask you before a destructive step, and mongrl’s instructions tell the assistant to as well.

It also comes with prompts for common jobs: Review listings waiting for approval, Sales summary, Orders that need us, Answer support conversations and Improve a listing.

Things to ask

Which listings are waiting for approval? Recommend what to do with each, but don't change anything yet.

How did we do in September compared with August? Sales, our commission and payouts, in kronor.

Find the orders we deliver that were paid more than three days ago and aren't marked delivered.

Draft replies to the unread support conversations, in each customer's language. Show me before sending.

For developers

• Streamable HTTP, stateless, with JSON responses. Authorization follows the MCP specification: a request without a token gets 401 with WWW-Authenticate naming /.well-known/oauth-protected-resource/mcp (RFC 9728), whose authorization server supports dynamic client registration (RFC 7591) and PKCE.
• Tools are the console’s endpoints, and their arguments are the endpoints’ own, from the same spec as /openapi.json. Money is in minor units, as everywhere in mongrl.
• A failed call comes back as a tool error with the API’s problem detail and, for invalid arguments, which field it was, so the assistant can correct itself.
• Building a storefront rather than running a marketplace? See Build with an AI agent.