Pattern: Risk-Tiered Access
Summary
MCP tools are grouped into risk tiers. Low-risk read-only tools are freely available to all agents. Higher-risk write and admin tools require additional authentication, human approval, or are blocked entirely for automated agents.
Tiers
| Tier | Examples | Auth Required |
|---|---|---|
| Read | Search docs, query status | API key |
| Write | Create ticket, send message | OAuth + scope |
| Admin | Deploy code, delete records | MFA + human approval |
When to use
- When agents need write access but blast radius must be controlled
- Regulated environments with data classification requirements
- When different agent types (human-in-the-loop vs. fully automated) need different permissions
Real-world reference
Linear and Cloudflare both implement risk-tiered MCP access in their published server implementations.