Tool permissions & policy
An agent can only do what its tool list allows. Policy means deciding what goes on that list. Allow the core coding loop, deny the tools that reach outside until you need them, and hold the irreversible ones for a person to approve.
These posts cover how to write an allowlist, which tools to deny by default, and why per-call permissions beat roles assigned at login. They also cover how rules inside the agent get bypassed, and what enforcement outside the agent process looks like: a model proxy, gateway-executed tools and an OS sandbox. Which of those you get depends on how you launch the agent. For Claude Code, the model proxy only applies when you start sessions with the claude-acp launcher. Plain claude keeps the hook only.
The Tool Surface Index lists every tool Claude Code and Codex declare, grouped by blast radius. The incident-by-incident version is the Stop Your Agent From… series.
Posts on tool policy
-
AI Agent Tool Allowlists: Deny by Default, Scope per Task, Audit Everything
One Claude Code session declared 76 tools and 64 never fired. A tool allowlist is the list of calls your agent may make, with everything else denied. How to set one in Claude Code, Codex CLI and MCP, and where client-side lists stop holding.
-
Claude Code Permissions: Which Tools to Deny or Ask About by Default
One Claude Code session declared 76 tools and used 12. The coding loop isn't where the risk is. The tools you never see it call are: send, schedule, publish, spawn. A default setup, argued from what each one can do.
-
Is --dangerously-skip-permissions Safe? What the Flag Turns Off
The flag drops Claude Code's permission prompts. It does not drop your deny rules, your PreToolUse hooks, or the rm -rf circuit breaker. Here is what still holds, what really turns the controls off, and what to set up before you use it.
-
Runtime authorization for AI agents: per-call permissions
Runtime authorization checks policy on every tool call an agent makes, not once at login. Why role-based access breaks at agent speed, and what per-call checks do.
-
How to stop a coding agent from dropping your production database
If your agent can reach a destructive credential, it can use it. Here is how to hold destructive commands for approval, shut them off for unattended runs, and keep production tokens out of the agent's reach.
Related: Claude Code · MCP security · how ACP policies work · all posts