If your company is going through a SOC 2 audit and someone on the team has connected Claude or Cursor to a handful of MCP servers, the auditor is going to ask about it — even if nobody thought of it as part of the “system” being audited. The moment a model can call a tool that reads or writes customer data, that tool call is inside the boundary SOC 2 cares about. This post maps the Trust Services Criteria an auditor actually tests onto an MCP deployment, so you know what evidence to have ready before the request comes in.
Why MCP ends up in scope at all
SOC 2 doesn’t name MCP, agents, or AI clients. It asks whether you have controls over who can access systems holding customer data, whether changes to those systems are tracked, and whether you’d notice if something went wrong. An MCP connection adds a new way in: a config file on someone’s laptop, or a gateway, that lets an AI client call tools against a real backend. From an auditor’s seat that’s a new access path to systems already in scope, provisioned outside the identity system the rest of your access reviews cover. It gets tested the same way any other access path does.
The criteria that actually touch an MCP setup
CC6 — logical access. The tester wants to see that access is provisioned per person, reviewed, and revocable in one action. A shared API key for a server, pasted into everyone’s claude_desktop_config.json or .cursor/mcp.json, fails this outright: nobody can prove who used it, and removing one person means rotating a key for the whole team. Per-user OAuth with scoped grants is what satisfies this control; a static secret in a plaintext file is the opposite of it.
CC7 — system operations and monitoring. The tester asks how you’d detect unauthorized or anomalous activity. “We’d see it in the model’s chat transcript” is not an answer a tester accepts, because transcripts live on individual machines and nobody reviews them. A central, queryable log of every tool call — who, which client, which server, which tool, the outcome — is. What that log needs to contain is covered in the MCP observability post.
CC9 — risk mitigation and vendor management. Every MCP server you connect is a vendor the moment it starts handling your data, whether or not anyone ran it through vendor review. The tester wants a list of what’s connected, evidence that each one was checked before approval, and a process for re-checking it. That’s a private registry or allow-list backed by an actual vetting step — see how to vet an MCP server — not a folder of config files nobody tracks.
Change management. MCP servers can change their own tool definitions between sessions with no release process on your end at all — a server approved for read-only search today can quietly add a write tool tomorrow. A tester who understands this will ask how you’d catch that. “We re-scan servers on an interval and alert on drift” is the answer; see rug pulls and drift detection.
What evidence actually satisfies a tester
Testers sample. Expect to be asked for a specific example, not a policy document:
- Pick one departed employee and show their MCP access was revoked the same day, not “eventually, when someone remembers.”
- Pull one real tool call from the log — timestamp, user, client, server, tool, arguments, outcome — and show it matches what actually happened.
- Walk through what happens if a connected server is suspected of being compromised: who gets notified, what gets cut off first, what gets checked. See what to do when an MCP server might be compromised for the actual sequence.
If any of those three turns into “let me check with the engineer who set that up,” that’s the gap the audit will flag.
A sequence that works before the audit, not during it
- Inventory every MCP server anyone in the company has connected, personal laptops included.
- Replace shared keys with per-user OAuth wherever the server supports it.
- Route every connection through one point that logs calls centrally, not per-device.
- Put new servers through a vetting step before approval, and re-check approved ones on a schedule.
- Write down — and actually test — what happens when a server is suspected of being compromised.
None of this is MCP-specific in spirit; it’s the same logical-access and monitoring story SOC 2 has always told. What’s new is that the access path is a model calling a tool it chose, not a person clicking a button, and most access-review tooling was never built to see it. A gateway is what gives that access path the same visibility your other systems already have.