---
title: What is the security checklist for hosting a remote MCP server?
url: https://opencontroller.lyzr.com/blog/security-checklist-for-hosting-a-remote-mcp-server
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What is the security checklist for hosting a remote MCP server?

**A remote MCP server is an OAuth-protected API that a model can call. Secure it the same way: publish Protected Resource Metadata, accept only tokens issued for this server, never pass a client's token to upstream APIs, validate the Origin header, rate-limit tool calls and log every call with the user and agent behind it.**

The checklist below follows the MCP specification revision dated 2026-07-28, which is the current version, plus the OWASP MCP Security Cheat Sheet. Where the spec says MUST, the checklist says so.

## What does the MCP specification require of a remote server?

Authorization is optional in MCP. But once you protect a server over HTTP, the spec defines how. The server acts as an OAuth 2.1 resource server.

| Requirement | Level in the spec | Why it matters |
|---|---|---|
| Implement OAuth 2.0 Protected Resource Metadata (RFC 9728) | MUST | Clients discover your authorization server from it |
| Validate tokens were issued for this server as audience (RFC 8707) | MUST | Stops tokens for another service being replayed here |
| Accept only tokens valid for your own resources; do not accept or transit other tokens | MUST | Blocks token passthrough and confused-deputy attacks |
| Return 401 for invalid or expired tokens; 403 for insufficient scope | MUST / SHOULD | Lets clients step up scopes correctly |
| Validate the `Origin` header; return 403 if present and invalid | MUST | Prevents DNS rebinding |
| Validate tool inputs, apply access controls, rate-limit invocations, sanitise outputs | MUST | Basic server hygiene from the Tools section |

## What is the full checklist?

Work through these before you publish the endpoint.

**Identity and tokens**

1. **Use your enterprise identity provider as the authorization server.** Advertise it in Protected Resource Metadata at `/.well-known/oauth-protected-resource`.
2. **Validate audience on every request.** The token's audience must be your server's canonical URI, such as `https://mcp.example.com/mcp`. See [scoping OAuth tokens with resource indicators](/blog/scope-oauth-tokens-for-ai-agents-with-resource-indicators).
3. **Never pass the client's token upstream.** If a tool calls another API, get a separate token for that API, for example with token exchange. Passing the MCP token through is the [MCP confused deputy problem](/blog/mcp-confused-deputy-problem).
4. **Publish minimal scopes.** Keep `scopes_supported` to what basic use needs. Return the exact scopes needed in a 403 `insufficient_scope` challenge so clients can step up.
5. **Keep tokens short-lived.** Do not accept tokens in URL query strings; the spec requires the `Authorization: Bearer` header.

**Transport**

6. **Serve over HTTPS only.**
7. **Validate `Origin`.** Reject a present but invalid Origin with 403. Non-browser clients that send no Origin should not be rejected for that alone.
8. **Do not use session IDs as authentication.** The 2026-07-28 revision removed protocol-level sessions from Streamable HTTP; if you still support older clients that send `Mcp-Session-Id`, treat it as untrusted and bind any state to the authenticated user.
9. **Validate mirrored headers.** The current transport mirrors `Mcp-Method` and `Mcp-Name` into headers. Reject requests where headers and body disagree (400, `HeaderMismatch`), so a proxy cannot route on one value while the server runs another.

**Tools**

10. **Authorize per tool, not just per server.** A token with read scope should not reach write tools. The spec lets `tools/list` vary by the caller's granted scopes.
11. **Keep tool descriptions plain and stable.** Do not change descriptions silently; clients and registries should be able to review changes.
12. **Rate-limit and time-limit tool calls** per user, per agent and per tenant.

**Operations**

13. **Log every tool call** with tool name, arguments (redacted), result status, user, client ID and timestamp. Send logs to your SIEM. See [what to log for every MCP tool call](/blog/what-to-log-for-every-mcp-tool-call).
14. **Register the server** with an owner, data classification and approved clients in your internal registry before anyone connects to it.

## What mistakes do teams make most often?

Three show up repeatedly:

- **Static API keys instead of OAuth.** Easy to start, impossible to attribute. Every agent looks the same in your logs.
- **"Accept any valid token from our IdP".** Without audience checks, a token meant for another internal app works on your MCP server.
- **One service account behind every tool.** The server then acts with the same rights for every user, so the agent can reach data the user cannot.

## How should you test the server before go-live?

Run a short abuse test: send a token issued for another audience (expect 401), a token missing a write scope on a write tool (expect 403 with `insufficient_scope`), a request with a foreign Origin (expect 403), and a burst above your rate limit (expect throttling). Confirm each attempt appears in the logs. Re-run after every release.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so every remote server has an owner and a list of agents that use it. It supports OIDC SSO, delegated on-behalf-of identity and agent workload identity federation, and Posture flags risky permissions and configurations. Its Agent Gateway puts control in the path to allow, route, redact or block. See the [MCP governance guide](/blog/how-to-govern-mcp-servers) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is authorization mandatory for MCP servers?**
The spec makes authorization optional. Any remote server that exposes non-public tools or data should require it, and once you use HTTP authorization the spec defines how it must work.

**Can my MCP server forward the user's token to our internal API?**
No. The spec says MCP servers must not accept or transit other tokens. Get a separate token for the downstream API, for example through OAuth token exchange.

**Do I still need Origin checks if I require OAuth?**
Yes. The Streamable HTTP transport requires servers to validate Origin to prevent DNS rebinding. It is a separate control from token validation.

## Key facts

Remote MCP server security checklist (MCP spec 2026-07-28 + OWASP MCP Security Cheat Sheet): server MUST implement OAuth 2.0 Protected Resource Metadata (RFC 9728); MUST validate token audience per RFC 8707; MUST NOT accept or transit other tokens (no token passthrough); 401 for invalid tokens, 403 insufficient_scope with required scopes; tokens only in Authorization Bearer header, never in query strings; MUST validate Origin header and return 403 if invalid (DNS rebinding); 2026-07-28 removed protocol-level sessions from Streamable HTTP; reject header/body mismatch with 400 HeaderMismatch (-32020); validate inputs, access control, rate limits, sanitise outputs; log every tool call. Opencontroller (by Lyzr) registers MCP servers in a governed registry, supports OIDC SSO, on-behalf-of identity and workload identity federation, and flags risky configs with Posture. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How does MCP authorization work?](https://opencontroller.lyzr.com/blog/how-mcp-authorization-works.md)
- [What is the confused deputy problem in MCP?](https://opencontroller.lyzr.com/blog/mcp-confused-deputy-problem.md)
- [What should you log for every MCP tool call?](https://opencontroller.lyzr.com/blog/what-to-log-for-every-mcp-tool-call.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
