DEV Community

Cover image for Who Controls What an AI Agent Can Ask? Governance in an MCP World
Harshit Chouhan
Harshit Chouhan

Posted on Originally published at colrows.com

Who Controls What an AI Agent Can Ask? Governance in an MCP World

Your MCP server exposes tools. Your agent calls them. Nobody in that sentence checked whether it should.

Most MCP deployments have no answer to "who authorised this question?"

Tool-level permissions aren't enough

The usual first attempt is binary: this agent may call query_warehouse. Business data isn't binary. The same tool needs to return different rows depending on who is behind the request.

That means identity has to survive the whole chain — user → agent → MCP server → query — and most implementations drop it at the server boundary, where the connection has its own service credentials.

What governance actually requires

Requirement Why it matters
Identity propagates to the query Otherwise every caller is the service account
Entitlement evaluated per question Not per tool, not per connection
Ambiguous or unauthorised intent fails to compile Not filtered afterwards
Audit ties person → intent → SQL → result The only defensible artefact

The third row is the architectural one. If your answer to an unauthorised request is "we filter the result," the data already left the warehouse and you're describing a breach with extra steps.

Where to put it

Not in the agent — agents are the thing being governed. Not in the prompt — prompts are suggestions. In the layer that compiles the query.

Attach policy to business concepts rather than tables, evaluate it during context resolution, and inject RBAC, ABAC and row/column predicates before SQL is emitted. Then MCP governance stops being a policy document and becomes a property of the system.

The protocol is settled. What sits behind it is still your decision.


The full breakdown — the identity propagation model, policy composition across scopes, and the audit format — is here:

👉 Who Controls What an AI Agent Can Ask? Governance in an MCP World


Originally published at colrows.com/blogs/mcp-governance

Top comments (0)