Safeguard
AI Security•

How do you answer 'Do you inventory your MCP servers?'

Reviewers ask whether you know which MCP servers your agents can reach. Here is a truthful answer, the fields to track, and the evidence to offer.

Safeguard Security Team
3 min read

An honest answer states whether you keep a register of MCP servers that company agents and assistants can connect to, what it records for each one, who approves additions, and how often it is reviewed. If the register is partial, say which scope it covers and when the rest will be done.

Why this question appears now

Model Context Protocol servers give an assistant tools, such as reading files or calling an API. The MCP specification says local servers may have direct access to the user's system, and that clients should warn that servers run with the same privileges as the client. A reviewer who reads that sentence will want to know how many such servers exist in your company.

Inventory is the cheapest control that makes the others possible. You cannot gate, scope or revoke what you have not listed.

What the register should hold

FieldWhy it matters
Server name and sourceWhere the code comes from
Local or remoteLocal servers run on the laptop with its privileges
Transport and endpointstdio or HTTP, and which host
Tools exposedRead only or able to change things
Credentials usedWhich token, with what scope
Owner and approverA person, not a team alias
Approval date and review dateShows it is maintained
Client apps using itEditors, desktop apps, agents

Sample wording

"Yes. We keep a register of MCP servers approved for use with company assistants. Each entry records the source, whether it runs locally or remotely, the tools it exposes, the credential and its scope, an owner and a review date. Additions need security approval. The register covers engineering laptops and our CI agents; coverage of the data team is scheduled for next quarter."

The last sentence matters. Stating the boundary is more credible than claiming total coverage.

Where partial answers go wrong

  • Counting only servers someone submitted. People install them from a project config file without telling anyone.
  • Listing a server without the credential it uses. The token is the real risk.
  • Recording a version once and never checking. Tool lists can change after approval, so note the review cadence.

A discovery step helps. Search repositories for MCP config files, review editor settings during onboarding, and compare against the register monthly. See also how to register MCP servers and agents in an inventory.

Evidence to offer

A redacted export of the register, the approval procedure, and the date of the last review. Do not send raw tokens or internal hostnames.

Safeguard offers asset discovery and agent identity for this kind of inventory, and an MCP Guardrail for controlling which tools agents can call.

Try it yourself

Click through it at your own pace, or open it full screen. Notice which details you would need to record for each server you connect.

Quick check

Does a remote MCP server belong in the inventory?

Yes. A remote server is a third party that receives your agent's requests, so it needs an owner, a credential record and a data handling note.

How often should the register be reviewed?

Pick a cadence you will keep, such as quarterly, and state it. A review you can prove beats a monthly promise you cannot.

Sources

Last reviewed: 19 October 2026.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.