docs: add security notes for network-exposed MCP servers

This commit is contained in:
Theodor N. Engøy 2026-02-07 12:55:02 +01:00
commit 2975f605c8
2 changed files with 34 additions and 0 deletions

View file

@ -3,3 +3,7 @@
A repository of servers and clients from the following Model Context Protocol tutorials:
- [Quickstart](https://modelcontextprotocol.io/quickstart) – a simple MCP weather server
- [Building MCP clients](https://modelcontextprotocol.io/tutorials/building-a-client) – an LLM-powered chatbot MCP client
## Security Note
These examples are intentionally minimal. If you expose an MCP server over a network (HTTP/SSE/WebSocket), add authentication and basic hardening (CORS allowlist, request size limits, timeouts, rate limits, and log redaction). See [`SECURITY.md`](./SECURITY.md).

30
SECURITY.md Normal file
View file

@ -0,0 +1,30 @@
# Security Notes (Quickstart Resources)
These examples are intentionally minimal and optimized for learning. Before using similar patterns in production, apply basic hardening.
## If You Expose An MCP Server Over A Network (HTTP/SSE/WebSocket)
- **Require authentication**: do not expose unauthenticated tool endpoints to the public internet.
- **Treat browsers as hostile**:
- Do not use wildcard CORS (`Access-Control-Allow-Origin: *`) on authenticated endpoints.
- Do not reflect `Origin` without allowlist validation.
- Prefer an explicit origin allowlist.
- **Bound resource usage**:
- Set an explicit maximum request body size.
- Add timeouts to outbound requests.
- Add rate limits (per user/token and/or per IP).
- **Avoid RCE primitives in tools**:
- Avoid `eval` / dynamic code execution.
- Avoid invoking a shell with attacker-controlled input (`exec`, `sh -c`, `shell=True`).
- If you must run commands, enforce strict allowlists and pass arguments as arrays (no shell).
- **Don't leak secrets**: redact `Authorization`, cookies, and API keys from logs.
## Local-Only Usage
If you run these examples locally via stdio transports, your main risks are still:
- accidentally adding dangerous tools (filesystem/shell) without strict controls
- leaking secrets via logs or environment
When in doubt: keep tool capability narrow and add input validation.