Tutorials
Connect an agent over MCP
Give Claude Code, Cursor, or any MCP client read and write access to a COUNT workspace — one command to add the server, an OAuth consent screen to scope it, and then plain questions instead of endpoints.
4 steps · about 34 seconds · from one command to an answer out of the books
What you just saw
- 1The remote server needs no API keyThe transport is HTTP and the auth is OAuth, so the config holds a URL and nothing secret. VS Code and Cursor both take a one-click install of that same config.
- 2Consent sets the ceilingThe workspaces ticked on the authorize screen are the only ones the agent can reach, for as long as the token lives. Revoking the connection in COUNT cuts it off immediately.
- 3Tools are grouped like the API referenceA tool name tells you which endpoint it reaches, and read-only resources let an agent pull context — the chart of accounts, customers, vendors — without spending a tool call.
- 4The agent resolves the workspace before it readsA token can cover several workspaces, and they are separate books. A well-behaved client picks one rather than blending them into a single answer.