Is API key auth enough?

It is a start, but production should also use scopes, rotation, logging, and approval policies.

Can MCP leak secrets?

Yes if tools return secrets or logs store raw headers. Redact aggressively.

Should tools run in sandboxes?

Use sandboxing for file, shell, browser, or network-capable tools.

complianceHowTo

MCP Security Best Practices

Secure MCP servers with auth, least privilege, tool approval, sandboxing, rate limits, and safe logging.

Quick Answer / TL;DR

Secure MCP by denying risky tools by default, validating schemas, isolating secrets, rate-limiting calls, logging safely, and requiring human approval for destructive actions.

Key Takeaways

Least privilege first.
Never trust tool arguments.
Mask secrets in logs.

Baseline policy

Security starts with a clear boundary. Every server should know which resources it may read, which actions it may perform, and which users or clients may invoke it.

{
  "auth": "required",
  "rateLimit": "60/minute",
  "sandbox": true,
  "denyTools": ["shell.exec", "payments.refund"],
  "requireApproval": ["write", "delete", "send_money"]
}