MCP Security Checklist for AI Agent Teams
MCP makes tool access cleaner.
It does not make tool access safe by default.
If an MCP server can read files, hit internal APIs, call SaaS tools, or execute local commands, you need a security checklist before the agent gets near it.
1. Know the server owner#
For every MCP server:
- Who maintains it?
- Where is the source?
- How is it deployed?
- How are updates reviewed?
- What tools does it expose?
- What data can it access?
Unknown MCP servers are untrusted code.
2. Separate read and write#
Do not mix all capabilities together.
Split:
- Read resources.
- Read tools.
- Write tools.
- Admin tools.
- Dangerous tools.
The agent should not get write access because it needed search.
3. Defend against prompt injection#
Retrieved content can be hostile.
Treat docs, issues, web pages, tickets, and comments as untrusted unless proven otherwise.
The policy layer should prevent content from granting itself authority:
- Docs cannot approve actions.
- Tickets cannot change tool policy.
- Web pages cannot request secrets.
- Repo files cannot override workflow permissions.
4. Keep secrets out of the model#
The MCP server can use credentials.
The model should not see them.
Return facts, not raw secrets. Redact by default. Log what was redacted.
5. Sandbox risky tools#
Browser automation, shell execution, code execution, and file access need stronger boundaries.
Use:
- Workspace isolation.
- Read-only mounts.
- Network limits.
- Timeouts.
- Resource limits.
- Human approval for destructive actions.
6. Log every meaningful call#
Audit:
- Agent.
- User.
- Workspace.
- Server.
- Tool.
- Arguments.
- Scope.
- Approval.
- Result summary.
- Error state.
If a tool can affect production, it needs a trace.
Build it in Codelit#
Try this:
Create an MCP security checklist for an AI agent team. Include server trust, read/write separation, tool permissions, prompt injection defense, secret handling, sandboxing, audit logs, approval gates, and rollout.
Build the MCP security checklist
MCP is a powerful boundary. Treat it like one.
Continue learning
Go deeper on AI agents and LLM systems
As an Amazon Associate I earn from qualifying purchases. Codelit may receive a commission at no extra cost to you.
AI Engineering
Chip Huyen · 2025
Book-length treatment of evals, prompts and deployment tradeoffs for foundation-model apps.
AI Agents in Action, Second Edition
Micheal Lanham · 2026
Working code for agent memory, tools, MCP and A2A, in the July 2026 second edition.
4.8 (29)Kindle editionModel Context Protocol for LLMs
Naveen Krishnan · 2026
436 pages on MCP servers, gateways and auth patterns, past what the spec docs explain.
4.4 (12)API Security in Action
Neil Madden · 2020
Working code for OAuth2 scopes, macaroons and capability URIs, with the failure modes of each.
4.7 (48)Kindle edition