The Model Context Protocol is an open standard that lets AI clients connect to external systems through a uniform interface, which means a compliance platform can expose its live control, evidence and risk data to tools like Claude, ChatGPT or Cursor without a bespoke integration per client. The practical gain is that questions people already ask — which controls are failing, what evidence expires this month, which vendors touch production data — get answered from current state rather than a spreadsheet someone exported in March. The questions to settle before connecting one are scope, credentials and audit: what the key can read, who holds it, and whether you can reconstruct afterwards what was accessed.
What the protocol actually is
The Model Context Protocol is an open standard for connecting AI clients to external tools and data sources through a uniform interface. Before it, every combination of assistant and system needed its own integration. With it, a system exposes one server and any client that speaks the protocol can use it.
For compliance data that matters more than it might sound, because the audience for this data is unusually varied. An engineer wants to know why a control is failing while they are in their editor. A compliance manager wants a summary in a chat client. A security lead wants to ask about vendor exposure without learning a new dashboard. One server serves all three.
The questions that get better
The gain is not that a model can phrase an answer nicely. It is that the answer reflects current state. Compare asking about overdue evidence against a quarterly export with asking the same question against the live programme.
- Which controls are failing right now, and since when.
- What evidence expires in the next thirty days, and who owns it.
- Which vendors have access to production systems or personal data.
- What changed in access control since the last review.
- Where are we blocked for the framework we are pursuing, ranked by what unblocks the most.
Each of these is answerable today by someone who knows where to look. What changes is the cost of asking, which is what determines whether the question gets asked at all.
Scoping the key
An MCP server for compliance data is guarded by an API key, and the scoping decision deserves more thought than it usually gets. A compliance programme is a precise description of your weaknesses: which controls fail, which evidence is stale, which vendors hold what. That is useful to you and useful to an attacker in roughly equal measure.
- Issue keys per use case rather than one organisational key that ends up in several places.
- Grant read scopes only, unless a workflow genuinely requires writing back, and never grant write access for exploratory use.
- Treat the key as a production credential — rotate it, store it in your secret manager, and never paste it into a shared configuration file.
- Confirm what the client does with retrieved data, particularly whether it is retained or used for training, before connecting anything sensitive.
Audit and attribution
Whatever you connect, you should be able to answer afterwards what was accessed and by which key. This is the same expectation you would apply to any other integration reading sensitive data, and it is worth confirming it exists before you rely on the connection rather than after an incident.
The ISO 27001 access control and logging expectations apply here in the ordinary way: an MCP server is a system component with access to sensitive data, and it belongs in your asset inventory, your access reviews and your logging.
What a model answer is and is not
An assistant reading your compliance data produces a draft. It can be wrong about scope, it can summarise selectively, and it will sound equally confident either way. That is fine for the questions it is good at — orientation, triage, drafting a narrative from evidence that already exists — and unsuitable for anything you would put in front of an auditor without checking.
The useful framing is the same one that applies to AI anywhere in a compliance programme: the model accelerates finding and drafting, and a person remains accountable for the conclusion. An assistant that says your controls look healthy has not performed an assessment.
Getting started
Noru exposes an MCP server so AI clients can query your live compliance programme, with setup guides for the common clients under MCP integrations. Start with a read-only key scoped to one use case, ask the questions you already ask manually, and compare the answers against the dashboard before you trust them.