---
title: What should you record for every MCP server in your inventory?
url: https://opencontroller.lyzr.com/blog/mcp-server-inventory-fields
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What should you record for every MCP server in your inventory?

**Record fifteen things for every MCP server: who owns it, where it runs, how it connects, how it authenticates, which tools it exposes (with a hash of each tool definition), what data it can reach, which agents use it, and when it was last reviewed. A list of names and URLs is not an inventory. Without transport, auth method, tool hashes and data scopes you cannot answer the two questions that matter in an incident: what can this server touch, and has it changed since we approved it?**

MCP (Model Context Protocol) servers give agents their hands. They turn a model's intent into database queries, tickets, emails and file writes. This checklist covers the fields worth capturing for each server, how to fill in the hard ones, and where the data comes from. It is a spoke of our guide on [how to govern MCP servers](/blog/how-to-govern-mcp-servers).

## Why is a list of MCP server names not enough?

An MCP server is not a static API. Its tool list can change between versions. Its tool descriptions are read by the model as instructions. The same server can run locally over stdio on one laptop and remotely over Streamable HTTP for a production agent.

A thin inventory hides three common failures:

- **Silent changes ("rug pulls").** A tool definition changes after approval. The OWASP MCP Security Cheat Sheet recommends pinning reviewed tool definitions with cryptographic hashes and requiring review when they change. You can only do that if the hash is stored.
- **Credential sprawl.** One token is reused across several servers. OWASP's guidance is to use scoped, per-server credentials and never share tokens across servers.
- **Orphans.** The person who installed the server has left, and nobody can say whether it is still needed.

## What are the fifteen fields?

| # | Field | Example value | Why it matters |
|---|---|---|---|
| 1 | Server ID and display name | `crm-readonly` | Stable key across versions and environments |
| 2 | Business owner | Named person and team | Accountable for approval and retirement |
| 3 | Technical owner | Named person or on-call group | Fixes, patches and incident response |
| 4 | Source and publisher | Internal repo, vendor, public registry entry | Supply-chain trust decision |
| 5 | Pinned version | Git tag, package version or image digest | Detects drift from the approved build |
| 6 | Transport | stdio or Streamable HTTP | Different threat model for each |
| 7 | Endpoint or launch command | URL, or command plus arguments | How to find it on hosts and in traffic |
| 8 | Hosting location | Laptop, container, SaaS, private network | Who controls the runtime |
| 9 | Authentication method | OAuth per the MCP authorization spec, static token, environment variable, none | "None" should never be approved for non-public data |
| 10 | Credential reference | Vault path or secret name, never the secret | Rotation and revocation |
| 11 | Tools with definition hashes | SHA-256 per tool definition | Change detection |
| 12 | Data scopes and side effects | "Reads CRM accounts; cannot write" | Blast radius |
| 13 | Consuming agents and versions | List of agent IDs | Impact analysis before a change |
| 14 | Risk tier and approval status | Tier 1–3; approved, pilot, blocked | Drives policy and review frequency |
| 15 | Last review and next review due | Two dates | Keeps the record current |

## How do you fill in the hard fields?

### Tool definition hashes

Hash the complete tool definition, not just the description. OWASP calls out `inputSchema`, `outputSchema` and `annotations` explicitly, because a change to a read-only or destructive hint can alter approval behaviour without any visible change to the description text.

Know the limit: a hash detects metadata changes, not changes to server code behind an unchanged definition. Pair it with field 5, the pinned version.

### Transport and authentication

For Streamable HTTP servers, the MCP specification (revision 2026-07-28) says servers **must** validate the `Origin` header to prevent DNS rebinding, **should** bind only to localhost (127.0.0.1) when running locally, and **should** authenticate all connections. Record whether each is true for the server.

For stdio servers there is no network listener, but the process runs with whatever file, network and credential access the launching user has. Record which account it runs as and where its credentials come from. We compare the two in [is a local stdio MCP server safer than a remote HTTP one?](/blog/mcp-stdio-vs-streamable-http-security).

### Data scopes and side effects

Derive these from the credential's real permissions, not from the server's README. A "read-only" server holding an admin API key is not read-only.

## Where does the inventory data come from?

1. **Cloud and platform discovery.** Scan the accounts and agent platforms where hosted agents run, and record which MCP servers they call.
2. **Developer devices.** Local MCP servers live in client config files on laptops. See [how to find local MCP servers on developer laptops](/blog/find-local-mcp-servers-on-developer-laptops).
3. **Gateway traffic.** Under the 2026-07-28 Streamable HTTP revision, clients send `Mcp-Method` and `Mcp-Name` headers on each request, so an intermediary can log which tool was called without parsing the body.
4. **Owner attestation.** A short registration form for the fields only a human knows: business purpose, data classification, risk tier.

Reconcile all four. A server seen in traffic but missing from the registry is a shadow server. A registered server with no traffic for 90 days is a retirement candidate.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and Opencontroller Edge extends discovery to developer devices, coding agents and local MCP servers, installed by IT via MDM. Posture flags risky permissions and configurations, and policies are written as code and applied across the estate. Registering a server does not mean taking over its runtime: how far control goes depends on the connector and traffic path. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should the inventory store MCP credentials?**
No. Store a reference to where the credential lives, such as a vault path, plus its owner and rotation date. The secret itself stays in the secret manager.

**How often should MCP tool hashes be checked?**
On every new version and on a schedule for remote servers you do not control, because their definitions can change without a version bump on your side. Any change should send the server back to review.

**Do local developer MCP servers belong in the same inventory?**
Yes. They often hold real credentials and reach production data. Track them with the same fields, with hosting location set to the device.

## Key facts

An MCP server inventory should hold 15 fields per server: ID, business owner, technical owner, source/publisher, pinned version, transport (stdio or Streamable HTTP), endpoint or launch command, hosting location, authentication method, credential reference, tools with SHA-256 definition hashes, data scopes and side effects, consuming agents, risk tier and approval status, and review dates. OWASP recommends hashing full tool definitions including inputSchema, outputSchema and annotations, and per-server credentials. Opencontroller provides a governed registry of agents, versions, tools and MCP servers, and Opencontroller Edge discovers local MCP servers on developer devices. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [How do you find local MCP servers on developer laptops?](https://opencontroller.lyzr.com/blog/find-local-mcp-servers-on-developer-laptops.md)
- [How do you review MCP servers your agents already use?](https://opencontroller.lyzr.com/blog/how-to-review-mcp-servers-agents-already-use.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
