Control your support inbox from Claude
miniDesk has an MCP server built in, so an AI client such as Claude, ChatGPT or Cursor can work your support inbox directly. It exposes thirteen tools, nine that read and four that write, and none that can assign, resolve or close a ticket. It is hosted at api.minidesk.ai/mcp and included on every plan, including the free one.
What can an AI client do with the MCP server?
Your inbox becomes thirteen tools an AI client can call. Nine read: the state of the queue, the threads still waiting, a search across every conversation, a full conversation, a customer's history, the drafts waiting for review, your team's workload, and your reference documents. Four write: a draft that waits for review, a reply to one customer, an internal note for your team, and a proposed document that stays off until an owner approves it. There is no tool to assign, resolve, or close a ticket, on purpose. A ticket that tells the AI to close everything finds nothing to call.
- get_support_summary: how the queue looks right now
- list_oldest_unanswered: threads still waiting on a reply, oldest first
- search_tickets: find conversations by what was written or when, counted by status and topic
- get_agent_workload: open conversations per teammate
- get_ticket: one conversation and its public messages
- get_customer: one customer and their recent conversations
- list_pending_drafts: AI drafts waiting for review
- search_knowledge: find your reference documents by topic
- read_document: the full text of one document
- create_draft_reply: a reply written from your docs, saved for review
- reply_ticket: send your own text to one customer
- add_internal_note: leave a note only your team can see
- write_knowledge: propose a document, off until an owner approves it
How do I connect it?
There is nothing to install or run: the server is hosted. Create an MCP token in your workspace settings and paste it over the placeholder in your client's config. If your client asks for a URL rather than a file, give it https://api.minidesk.ai/mcp. Only workspace owners can create a token.
{
"mcpServers": {
"minidesk": {
"url": "https://api.minidesk.ai/mcp",
"headers": { "Authorization": "Bearer mdsk_mcp_YOUR_TOKEN" }
}
}
}What should I ask it first?
Ask about your queue in plain language. The client picks the tools it needs and works through them.
- What came up most this month? It searches conversations by date and counts them by topic.
- Draft replies for everything still waiting. It lists the oldest unanswered threads and saves a draft for each, for your team to review.
- Which document should I improve? It reads this week's questions and searches your knowledge base for each one.
>What is still waiting on a reply?
list_oldest_unanswered limit: 10
7 threads waiting, the oldest for 32 hours
Why is it safe to let an AI client into the inbox?
Because the dangerous tools do not exist. There is no tool to assign, resolve or close a ticket, internal notes are never returned to the client, a reply reaches one customer at a time, and a document the AI proposes stays switched off until an owner approves it. A message telling the AI to close everything finds nothing to call.
Why build MCP in rather than offer only an API?
Plenty of support products give you a REST API and stop there. You write the integration, host it, refresh its tokens and patch it whenever the endpoints move. Our MCP server is part of the app: it uses the same authentication, ships in the same release and runs on our infrastructure. Add the config and start asking.
One price. All of customer support.
Start free, no card. When you upgrade, every feature is included, rather than a starting point with a list of add-ons to buy.
