Skip to content
Agentic Control Plane

Claude Code Permissions: Which Tools to Deny or Ask About by Default

David Crowe David Crowe · · 7 min read
policy control claude-code tool-calling default-deny blast-radius
Share X HN LinkedIn

Short answer: allow the coding loop (Read, Grep, Glob, Edit, Write, Bash inside a sandbox). Deny the tools that reach outside your machine until you need them: SendMessage, PushNotification, Artifact (publishes web pages), CronCreate, CronDelete, ScheduleWakeup, and remote-spawn tools like RemoteTrigger. Allow network reads (WebFetch, WebSearch) but review them in the log. Ask a human before browser control and MCP connectors. The rest of this post explains why.

Open a Claude Code session and count the tools it holds. In one real session we captured, Claude Code declared 76 tools to the model. The agent used 12. Sixty-four were never called. All 76 were available on every turn, including tools that send email, publish public web pages, schedule future runs of the agent, and drive a browser.

Most advice about Claude Code permissions covers three tools: edit files, run shell commands, ask first or not. That’s the wrong end of the list. The decisions that matter are in the tail you’ve never watched it use.

Don’t deny the core loop

Deny by default and approve each call sounds right. For a coding agent’s core loop it backfires.

Read, Grep, Glob, Edit, Write and Bash are how the agent works at all. If it has to ask before reading a file, you get hundreds of prompts a session. You won’t weigh hundreds of prompts. You’ll click yes. Two days in, the one prompt that mattered, the odd Bash call or the write outside the repo, looks the same as the 400 you already approved. That’s approval fatigue, and it’s worse than no gate because it feels like control.

So the question isn’t “deny everything?” It’s which tools deserve friction. Sorting the 76 by what a call can do answers it.

The tools fall into families

Names differ by harness, but the capabilities group the same way. This is the surface from that session:

FamilyTools (examples)What a call can doOut of the box
Read-onlyRead, Grep, Glob, CronListNothing outward. Worst case, it reads a file you'd rather it didn't.ALLOW
Local writeEdit, Write, NotebookEditChanges your working tree. Shows in git diff and can be undone.ALLOW
ExecutionBashAnything your user account can do, in one call.ALLOW + sandbox
Network readWebFetch, WebSearchMostly inbound. The quiet risks are data leaving in a URL and prompt injection coming back in the page.WATCH allow, review the log
SendSendMessage, PushNotification, MCP email and chat toolsLeaves the machine and reaches a person. You can't unsend it.DENY until needed
PublishArtifact (publishes hosted web pages)Public, has a URL, outlives the session.DENY
ScheduleCronCreate, CronDelete, ScheduleWakeupThe agent sets up its own future runs.DENY
SpawnAgent, Task, Workflow, RemoteTriggerStarts more agents, each with its own tools, sometimes on other machines.WATCH local ok · deny remote
MCP serversbrowser control, email, calendar, driveWhatever the server can do, the agent can do. Often the biggest single grant in the session.ASK case by case

Two families are easy to miss if you think of Claude Code as an editor.

Schedule. An agent that can call CronCreate or ScheduleWakeup can arrange to run again after you close the terminal. Pair that with anything in the send family and you have an agent that emails people while nobody is looking. Every other family assumes you’re at the keyboard. This one is about the runs where you aren’t.

Spawn. A subagent is a new loop with its own copy of the tools, and RemoteTrigger-style tools start runs on machines you aren’t looking at. Whatever you set for the main session, delegation is where it carries over or leaks.

The setup, step by step

1. Allow the core coding loop. Read, Grep and Glob freely. Edit and Write inside the project. Bash with whatever sandbox you have. Bash can do anything, so the control there is a sandbox plus a log, not a prompt you’ll stop reading.

2. Deny the outward tools: send, schedule, publish, remote-spawn. A coding task almost never needs them, and when one fires unexpectedly the effect leaves your machine. A denied SendMessage costs nothing on a day you didn’t need it. On a day you do, allowing it is a one-line change. Turn each one on at the first real need, one tool at a time.

3. Ask a human for the rare, irreversible calls, if you have a way to answer. Publishing a page, sending the email, creating the cron job: these are deliberate and can’t be undone, so a prompt fits. It only works for tools you call a few times a week, where you read each prompt. If nobody can answer the prompt, deny is the better setting. A gate nobody answers is a deny with extra waiting.

4. Recount now and then. The surface you audited isn’t the one you’re running. Tools load lazily, MCP servers connect mid-session, and updates add families. In our captured traffic one session gained 21 tools partway through the day when deferred tools loaded. No prompt, no changelog, just a longer list in the next request. Check the count again when it changes.

How to see your own tool list

You don’t need special tooling to start. Claude Code sends its full tool list, names and schemas, in every API request. Anything that sees that traffic can read it: a logging proxy, a gateway, or mitmproxy on your own laptop. Dump one request body, list the tool names, sort them into the families above, and you have the declared surface, not the documented one. Compare it across sessions and you can see drift.

Agentic Control Plane does this for you. It logs the tool calls Claude Code makes through its hook and lets you set each tool to allow, ask or deny. One thing to know: a new workspace starts in audit mode, which records calls and holds only the safety floor’s calls (force-push, curl | sh and similar). Rules you set apply once you turn on enforcement.

curl -sf https://agenticcontrolplane.com/install.sh | bash

The setup works with any tooling or none: coding loop open, outward tools denied until needed, irreversible calls asked, and a count you recheck. The 76 tools aren’t the problem. Not knowing which 76 you have today is.

The tool-surface control table for a live Claude Code agent: every declared tool a row with Allow / Flag / Deny / Approval controls, invoked vs never-invoked status, and a one-click suggested posture — deny 10 never-invoked outward-blast-radius tools

We keep captures of coding-agent tool lists for Claude Code, Codex and others in the Tool Surface Index.

Running Claude Code? One command adds a hook. Each tool call is recorded, the session's model calls are priced, and a force-push, destructive SQL or curl | sh waits for your OK in the session.

macOS · Linux · WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

What the installer writes  ·  Claude Code guide  ·  free up to 5 agents

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog