MCP servers are third-party vendors: what compliance teams need to know
Author
Audrey
NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.
Date Published
This post makes one narrow argument: an MCP server is a third-party vendor. If your third-party risk management (TPRM) program isn't treating it like one, you have a category of relationship touching your data that never went through the front door. Below: what MCP is in plain terms, why a server counts as a vendor, and what the July 2026 spec revision changed for your audit trail.
What MCP is, in plain terms
The Model Context Protocol (MCP) is the connector standard that lets AI tools reach other systems. Anthropic open-sourced it in November 2024. Think of it as a universal adapter: an AI assistant can talk to your calendar, your files, your ticketing system or a vendor's database through one common protocol.
The pieces that do the connecting are called MCP servers. Your staff don't install "MCP." They turn on a connector in a tool they already use, and MCP is what lets that tool reach out and touch something.
Examiners won't ask about the protocol by name. They will ask about its consequences: what can this AI tool reach, who approved that, and where's the record?
Why an MCP server is a vendor
An MCP server is software, usually run by someone else, that you grant network access and a token so it can read or act on your data. Rewrite that sentence without "MCP" and it's the definition of a third-party service provider. That is the thing your TPRM program exists to diligence.
The problem is how MCP servers arrive. A traditional vendor comes through procurement: a contract, a review, a place on the vendor list. An MCP server arrives when someone adds a connector, often in minutes, often without compliance learning about it.
The spec secures the connection. It doesn't vet the company.
Under the current authorization spec, an MCP server that supports authorization acts as an OAuth 2.1 resource server. It "MUST validate that access tokens were issued specifically for them as the intended audience." Clients must request tokens bound to the specific server they mean to use (RFC 8707). That makes a token issued for one server hard to replay at another.
Two details matter.
- Authorization is optional. The spec says so in its first requirement: "Authorization is OPTIONAL for MCP implementations." Servers using HTTP "SHOULD" conform when they support it. Servers running locally over STDIO pull credentials from the environment instead. So "it uses MCP" tells you nothing about how a given server authenticates.
- None of it covers the company. OAuth makes the handshake trustworthy. It says nothing about whether the party running the server should have your data, what they keep, or whether they'd tell you about a breach. Those are vendor questions, and vendor questions belong to you.
What the July 2026 revision changed for your audit trail
MCP versions are dates. Per the versioning page, the identifier is "the last date backwards incompatible changes were made." The current revision is 2026-07-28. Its changelog lists what changed since 2025-11-25. Four items matter to a compliance file.
The session is gone
The revision removes protocol-level sessions, the Mcp-Session-Id header and the initialize handshake. Every request now carries its own protocol version and capabilities. Clients "SHOULD identify themselves on each request" and servers should identify themselves in each result.
That's good engineering. It also means the protocol no longer ties a run of AI actions into one episode. If you want that grouping in a record, your vendor has to build it, for example with the OpenTelemetry trace context the spec now documents.
Logging is deprecated
The spec's deprecated features registry lists Logging as deprecated in 2026-07-28. The migration path: "Log to stderr for stdio transports; use OpenTelemetry for observability."
MCP's logging feature was never your audit trail. It carried debug messages from a server to the app driving it. What changed is that the protocol now says in writing that observability lives outside it. Whether you can answer "show me what your AI touched" depends on what your vendor chose to record.
Client registration moves to a URL you can read
Dynamic Client Registration, where software enrolls itself on the spot, is deprecated in favor of Client ID Metadata Documents, where a client's identity is a URL it publishes. "Which software holds this token, and who publishes it?" becomes a link you can open. That helps with the agent identity problem.
The revision also adds an issuer check (RFC 9207): clients "MUST validate a present iss against the recorded issuer before redeeming the authorization code."
Deprecated isn't removed
Deprecated Roots, Sampling, Logging and Dynamic Client Registration keep working until the "first revision released on or after 2027-07-28." The feature lifecycle policy sets a twelve-month minimum, shortened to at least ninety days for an active security risk. Nobody needs a rip-and-replace yet.
Where regulators point
The OCC's November framework (OCC Bulletin 2026-54) is an examiner tool. It says plainly that it "does not establish supervisory expectations for banks." Still, it names this surface. Section 2.11 lists "connected tools, and non-human identities" among things that "create attack surfaces," and adds: "The breadth of agent permissions and tool access may be treated as part of the attack surface." Section 2.9 says banks "may define the scope of authority, identity, permitted tool use, data use and escalation limits for provider agents that interact with bank systems and data." More in our November regulation update.
Credit unions: NCUA "lacks the authority to examine technology service providers," per GAO-25-107197 (May 2025). GAO first recommended Congress fix that in 2015. No examiner is backstopping the company behind an MCP server. The diligence is yours. See our NCUA vendor post for the proportional approach.
RIAs and broker-dealers: Reg S-P requires written policies "reasonably designed to require oversight, including through due diligence and monitoring, of service providers" (17 CFR 248.30(a)(5)). Providers must notify you no later than 72 hours after becoming aware of a breach. You can't receive that notice from a vendor you never recorded as one. More in our Reg S-P post.
What to do this quarter
- Put "MCP server" in your definition of vendor. One sentence in your TPRM policy: any external service granted access to firm or customer data, including MCP servers and AI connectors, is a third-party service provider subject to this program.
- Inventory the connectors. The AI tools your people use usually authorize through Google Workspace or Microsoft Entra. The app list and the scopes each holds sit in your admin console. Reading those grants is your first MCP map.
- Scope the token, then diligence the party. For each connector: what can it reach, for how many people, since when, who runs it, and can you revoke it in one move? Proportionality decides how deep you go. A read-only calendar connector gets a short note. Anything touching member or client records gets real diligence.
- Add two lines to your AI vendor questionnaire. "Which MCP specification revision does your product implement?" and "What record exists of actions your product takes on our data, how long do you keep it, and can we receive it?" OpenTelemetry in the answer means a standard-format trail exists.
- Write down your reasoning. "We reviewed this connector, it's read-only and non-core, here's our basis" is a sentence you can hand an examiner. An unreviewed connector isn't.
What to watch out for
- "MCP-compliant" badges. Following the protocol certifies nothing about the company on the other end. With authorization optional, it doesn't even certify strong auth. Ask for the vendor's posture.
- "It's read-only, so it's fine." Read-only still means a third party reads your regulated data. It lowers the action risk. The diligence obligation stays.
- "Stateless means it forgets." The application behind the protocol keeps whatever it likes. Retention is a contract question.
- Banning MCP outright. It's becoming standard plumbing in tools your staff already use. A ban pushes connectors out of sight, which is the failure you're trying to prevent.
Where NobleCloak fits
NobleCloak governs what an AI agent may do with a tool. Policy marks each tool as allowed, needs approval, or denied. Approval goes to a named person or role, and one approval covers one action. The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when. That's the record the July revision left to your vendors. See human in the loop.
For the inventory step, our read-only scan of Google Workspace and Entra lists every AI app your people authorized, the data scopes it holds, who's using it, and since when. Coverage is limited to what SSO and OAuth can see, so a local MCP server that never touches your identity provider won't show up. Our team runs the first report with you. Mapping scanned apps to connectable tools is coming soon.
We don't run vendor due diligence for you or certify any MCP server. That judgment stays with your program.

