MCP and AI assistants

Current runs an MCP server, the protocol AI assistants use to read and act in other tools. It offers the same operations as the REST API, as tools, through the same checks.

1. 1
   Add the connector

   In Claude, open Settings, then Connectors, and choose Add custom connector. On a Team or Enterprise plan, an owner adds it for the organization first.
2. 2
   Name it and paste the address

   Name it Current and paste https://api.current.day/mcp as the address. Choose Add.
3. 3
   Connect

   Choose Connect. Claude opens Current: sign in as a tenant admin and approve. Claude can then search and read your tickets and companies while you chat.

Anything a person or a connected tool wrote comes back inside an envelope with untrusted_text and source, and every tool answer is marked as holding untrusted data. The server tells the assistant that such text is data, never instructions, so a ticket that says "ignore your instructions" is read as part of the ticket.

Write tools start off. A tool that changes a ticket is listed only for an API agent's key that holds tickets:write, and only once Current has switched agent changes on; until then it answers exactly like a tool that does not exist. An app connected with OAuth never gets a write tool. No tool emails a customer, with one exception: a close, which does what your tenant admin chose for agent closes.

An agent's change otherwise travels exactly like a person's. A ticket that started in a customer's Teams chat shows its new status in that chat, and a change to a ticket linked to your PSA reaches the PSA, whose own notification rules may tell the customer.

Every tool answer fits Current's limits: 100 rows in a list, 15,000 characters in a text and 256 KB in all. Anything cut is marked, never cut silently. When a page of results had to give up rows, next_cursor is null and page_cut says how many were kept; ask again with the same cursor and limit set to that number (see Limits).