Tool governance
Prompt injection should not drop the production database.
Masking personal data does not stop a bad tool call. Guardrails inspect the JSON-RPC method and arguments first.
Three controls
Read-only mode allows get, list, and search, and refuses create, update, and delete. Allowlists shrink a huge upstream server down to the tools an agent actually needs. Argument rules catch DROP, TRUNCATE, and wildcard deletes even when the tool name looks harmless.
- Read-only toggle on every plan
- Per-tool rules on Team
- Per-client rules on Agency
Where it sits
The check runs on the gateway in front of your MCP server. The model never receives the upstream URL, so it cannot route around the policy by calling the database directly.
Questions
Does read-only mode break enrichment workflows?+–
Turn it off for the tools that must write, and keep it on for everything else. Team and Agency plans can differ by tool and by client workspace.
Is this a substitute for database permissions?+–
No. The upstream role should already be least privilege. Guardrails are a second gate for agent mistakes and prompt injection.
See it on a payload you control.
The sandbox and the free tools run before you create an account.