Does the send gate block drafts?

No. Drafts save without sending, so tools like create_draft pass. Tools with send, reply, forward, publish, post, pay, refund and similar words in the name are gated.

The Operator Kit for Claude Code

Claude can draft it. You decide if it sends.

Q: How do I stop Claude Code from sending email without asking?

A rule in CLAUDE.md is advice the model usually follows. A PreToolUse hook is code that runs before the tool call and can deny it. The free send gate is a hook that denies tools and commands that look like sends, posts or payments until you approve that action by typing a code in your own terminal.

Q: Does a PreToolUse hook work in bypass permissions mode?

Anthropic's hooks guide says a hook that returns permissionDecision deny blocks the tool even in bypassPermissions mode or with --dangerously-skip-permissions. Settings that turn hooks off, such as disableAllHooks, turn the gate off too.

Q: Can Claude approve its own send?

The approval script reads a 4-digit code from the terminal a person is sitting at, and Claude's shell has no terminal. A process that can open its own pseudo-terminal could still type it, so the gate is a safeguard, not a sandbox.

Q: Who sells the Operator Kit?

Kyle, the author of the free send gate, sells it for USD 19. The gate itself is free and MIT licensed.

A hook that blocks Claude Code from sending email, posting, or paying until you approve that one action in your own terminal. The gate is free. The kit around it is USD 19.

What it isA Claude Code PreToolUse hook. It denies sends, posts and payments until you approve that one action.

Who it's forPeople who give Claude Code access to email, a CRM, payments or customer files, and will use a terminal.

Free vs paidThe gate is free (MIT) and stands on its own. The USD 19 kit, which I sell, adds seven rule packs, a CLAUDE.md template, a memory starter and a 34-page PDF playbook.

Buy the kit, USD 19 Get the free send gate

Card or PayPal at checkout. You land on the download page right after paying. 30-day refund: email your PayPal receipt. If all you want is the gate, the free download is enough. Needs Claude Code and Node 20+. Built and run on macOS. Independent project, not affiliated with Anthropic.

The Operator Kit for Claude Code: a hook that blocks email, posts and payments until you approve each one. 156 tests passed, 0 failed.

What it is

A starter setup for people who let Claude Code act for them: email, a CRM, payments, customer work. It sits around a send gate, a hook that blocks Claude Code from sending, posting or paying until you approve that action. The gate itself is free and open source (MIT). The kit is what I use around it: seven rule packs that extend the gate to common business tools, a CLAUDE.md template, a memory system, two skills, three short docs on delegating and getting unstuck, and a PDF playbook that explains why each piece exists.

It comes from a setup I use day to day. I'm a home performance manager at a Portland contractor, not a software engineer. Everything specific to that company has been removed.

Free: the send gate

USD 0

The hook (send-gate.mjs), approve-send.mjs, away-mode.mjs
156-case test suite (test-send-gate.mjs)
Plugin install or install-gate.mjs, plus the 2026-10-25 test transcript
MIT licensed, source on GitHub

Download the free gate

The Operator Kit

USD 19

Everything in the free gate
Seven rule packs in packs/ (email and calendar, phone and SMS, chat and social, payments and invoicing, CRM, scheduling and field service, publishing), applied with apply-pack.mjs and checked by 156 more tests in test-packs.mjs
CLAUDE.template.md, with the reason for each section
memory-starter/ (a one-fact-per-file convention and four example memories), the grill-me and dream skills, three docs on delegating, writing skills and getting unstuck
PLAYBOOK.pdf, 34 pages, with a step-by-step install for people new to terminals
install.mjs and uninstall.mjs with --dry-run, and an effort-nudge.mjs hook

Buy the kit, USD 19

Card or PayPal at checkout. 30-day refund: email your PayPal receipt.

The free gate already blocks the sends it knows about. What the USD 19 adds is coverage for more business tools through the packs, and the files and PDF for setting up the rest of your Claude Code configuration around it.

How the send gate works

A hook runs before every tool call. It looks for tools that send, post or pay: MCP tools with names like send_email or create_refund, shell commands that send mail or POST to mail, chat, SMS, CRM, accounting or payment APIs, and browser scripts that click a Send button.
If it finds one, it denies the call and tells Claude to show you the exact text and recipient and get your yes first.
You approve in your own terminal window, not through Claude: you run approve-send and type a 4-digit code it shows you. That writes a single-use token for the next matching call, which expires after 180 seconds by default. Claude's shell has no terminal, so it can't type the code.
Away mode blocks every gated call, approved or not, until you turn it off in your own terminal.

Anthropic's hooks guide says a hook that returns deny blocks the tool even in bypassPermissions mode. Guide: how to stop Claude Code from sending email, posts and payments.

Terminal output of the send gate test suite: 156 passed, 0 failed.

Who it's for

People who already use Claude Code from a terminal, or are willing to learn three commands, and who give it access to email, a CRM, payments or customer files.
People who want Claude's rules written down in files instead of re-explained every session.

Who it's not for

Anyone who doesn't want to open a terminal at all. This is not a no-code product.
Anyone looking for a big pile of skills. There are two here, chosen because they get used.
Anyone who needs a guarantee. The gate is a safeguard with known limits, listed below.

What I have and haven't tested

The test suites run the hook directly with sample inputs. Passing them means the named cases behave as written. It doesn't mean every tool in your setup is covered.
On 2026-10-25, Claude Code 2.1.310 was run headless with --dangerously-skip-permissions and the gate wired in. A gated command was denied. The approve script, run from Claude's shell, exited without writing anything. After a code was typed at a pseudo-terminal, the gated command went through once. The transcript is in the zip.
I haven't run it in the interactive screen yet. The playbook's chapter 2 ends with a two-minute check you run yourself. Please do it before you rely on the gate.

What it can't see

A plain mouse click on a Send button through a computer-use or click-by-element tool.
APIs on hosts that aren't in its list or in a pack. You can add hosts in a config file.
git push, package publishes and deploys, unless you apply the publishing pack. Some MCP tools with unusual names, until you add them.
Approval is a typed code, not a sandbox. A process that can run Python can open its own pseudo-terminal and type it, and a determined agent can dodge the text guard that refuses commands naming the approve script.
Words like merge, comment, schedule and invoice gate some harmless tools too. You allow them one at a time.
If the hook can't read its own input, it lets the call through so a bad payload doesn't break your session. If node is missing, the settings wiring blocks the call. Anything that turns hooks off from outside Claude Code (disableAllHooks, --bare, or you editing settings.json yourself) turns the gate off too. Claude's own edits to settings.json are blocked.

Install

unzip operator-kit-v1.0.zip && cd operator-kit
node install.mjs --dry-run     # shows what it will copy and wire; changes nothing
node install.mjs
node tests/test-send-gate.mjs  # 156 passed, 0 failed

The installer backs up your settings.json first and never overwrites a file you already have unless you pass --force. Start a new Claude Code session afterward so the hooks load.

Common questions about stopping Claude Code from sending

How do I stop Claude Code from sending email without asking?

A rule in CLAUDE.md is advice the model usually follows. A PreToolUse hook is code that runs before the tool call and can deny it. The send gate is a hook like that: it denies tools and shell commands that look like sends, posts or payments, and tells Claude to show you the exact text and recipient. You approve by typing a 4-digit code in your own terminal. Full guide.

Does a PreToolUse hook work in bypass permissions mode?

Anthropic's hooks guide says a hook that returns permissionDecision: "deny" blocks the tool even in bypassPermissions mode or with --dangerously-skip-permissions. I confirmed it once, headless, on Claude Code 2.1.310. Settings that turn hooks off (disableAllHooks, or starting with --bare) turn the gate off too.

Can Claude approve its own send?

The approval script reads the code from the terminal a person is sitting at, and Claude's shell has no terminal. A process that can open its own pseudo-terminal could still type it, and the hook's text guard can be dodged. It is a safeguard against mistakes and over-eager automation, not a sandbox.

Does it block drafts?

No. Drafts save without sending, so something like create_draft passes. Tools with send, reply, forward, publish, post, pay or refund in the name are gated.

Who sells the Operator Kit?

I do (Kyle). The gate is free and MIT licensed. The USD 19 kit is the setup I use around it.

Guides

Questions

How do I get the files after paying?

PayPal sends you back to the download page as soon as the payment goes through. If you closed the window before that, email the address below with your PayPal receipt and I'll send the link.

Refunds?

30 days. Email with your PayPal receipt.

License?

The four send gate files are MIT. The rest of the kit is for personal and internal business use by the buyer and their team. Please don't resell or redistribute it.

Windows?

Use WSL. Linux should behave the same as macOS, but I haven't run it there.

Buy the kit, USD 19 Get the free send gate