Note
Agents and MCP: giving AI tools without giving away the keys
The Model Context Protocol standardises how agents reach tools and data. Governance is still your job.
Nov 2026 · 1 min read
What MCP standardises
The Model Context Protocol defines how an AI application (the host, through an MCP client) connects to servers that expose tools, resources and prompts. One protocol replaces a custom integration per tool, which makes it easy to give an agent a lot of reach, quickly.
Governance you still need
- Least privilege per server. Each server gets narrowly scoped credentials: read-only database views, a ticketing token limited to one project.
- Separate reading from acting. Put tools with side effects behind their own server and route them through human approval.
- Treat tool output as data. Content returned by tools can carry prompt-injection attempts; it must never change the agent's instructions or permissions.
- Log every call: the tool, inputs, outputs and the identity the agent acted for, so any action can be reconstructed.
- Pin and review servers. Install servers from sources you trust, pin versions, and review what each one can reach.
- Budgets and rate limits. Cap calls and spend per task; agents can loop.
A simple rule
If you wouldn't give a new contractor the permission on day one, don't give it to an agent.