Let an AI assistant do the books, with a person still approving every action.
Copperline's customers ask Claude or ChatGPT to pay bills, categorize spend, chase invoices and post journal entries. The assistant reads their ledger through the Apideck MCP server and queues each change. Nothing writes until someone approves it in Copperline.
What should we get done?
This assistant is connected to Copperline. It can read the books and queue work, but nothing writes until someone approves it in Copperline.
Any MCP client works the same way: Claude, ChatGPT, Cursor or your own agent. The tables are MCP Apps the server returns with its results.
Set in Copperline. The assistant can read everything and queue only what's ticked.
The policy the team already follows, applied to the assistant. Flagged items never auto-approve; everything else waits for a tap.
How it fits together
Reads pass straight through. A write the assistant asks for is held by Copperline, not run, and goes to the ledger only once a person, or a rule they set, approves it. The guardrail lives in the server the assistant calls, not in its prompt, so the assistant can't be talked past it.