Security
Read tools, write tools, and keeping AI actions safe
Most questions only need data. When you want your assistant to act, the tool should say so — and you should approve it.
Engineering, SellerPort MCP · · 5 min read
Giving an assistant the ability to change a price or pause a campaign is useful, and it deserves more care than giving it the ability to read a report. The design principle is simple: make actions explicit, narrow, and visible.
Explicit
Every tool that changes live data is described to the assistant as a WRITE operation, intended for requests that ask for a change. Reporting questions never need them.
Narrow
Write tools do one thing each — set a price for one SKU, change one campaign’s budget, acknowledge one purchase order. There is no general-purpose “update anything” tool.
Visible
Most MCP clients ask for approval before running a tool, and every call appears in the query log. For teams that only report, issue separate MCP tokens and approve only read tools.