Read-only is a property of credentials, not of prompts
Telling a model not to change anything is a request. Giving it credentials that cannot change anything is a guarantee. Here is how to build the second.
RoutMe team
Product and engineering
A support assistant on a public website talks to anyone. Some of those people will try to talk it out of its instructions. If you rely on the prompt to keep it from changing your data, you are relying on the weakest layer in the stack.
Four layers, ranked by strength
- Credentials. The token the tool uses can only read. The system refuses writes no matter what the model asks for.
- Tools. Only read operations exist as tools. There is no refund function to call.
- Platform filtering. The platform only offers tools the server marks read-only. In RoutMe's chat assistant, a tool is offered only if the server marks it with the readOnlyHint annotation.
- Prompt. An instruction not to change anything. Useful, and the first thing an attacker targets.
A caveat about annotations
The MCP specification says clients must treat tool annotations as untrusted unless they come from trusted servers. A read-only annotation is a claim by the server author. That is fine when you run the server, and it is why the real guarantee is a read-scoped credential on that server.
On Shopify
A Shopify custom app token carries the scopes you grant. Read access to orders is the read_orders scope, which covers recent orders unless read_all_orders is also granted. Grant read scopes only, and keep the token on your MCP server, never in the page.
For the longer version, see what read-only AI means and what MCP is.
- MCP
- Safety
- Read-only
RoutMe team
Product and engineering
Practical guides from the people who build RoutMe, with every product claim checked against what is live today.
Keep reading
- Engineering
Connect your systems without glue code: MCP, APIs, databases
How RoutMe reaches your data without glue code: MCP servers, your own APIs and read-only database views. What each is, when to use it and how to keep it safe.
- Engineering
Giving a support agent your data without giving it your database
An agent that cannot read your systems can only apologise. Here is how to expose real data through MCP tools with a boundary you can actually defend in a security review.
See it on your own data
Connect a system, drop an agent node into a flow, and watch it answer from what is actually true in your systems right now.