Claude Code Function Hooks (Experimental): How to Enable Them and Write Your First TypeScript Hook

What Function Hooks Are

Today a Claude Code hook is a shell command. Claude Code sends it a JSON payload on stdin, the script exits 0 or 2, and optionally prints JSON to allow, deny, or add context. It works, but the contract is narrow: a shell hook cannot rewrite a tool's input, cannot return its own result in place of the tool, has no state between calls, and cannot draw anything on screen.

The proposal adds a fifth hook type beside command, prompt, agent and http: a function. Your hooks.json names a TypeScript module, the module exports a register function, and every hook inside it has the same signature: ($, e, next). If you have written Express or Koa middleware, you already know the shape.

How to Enable It

The videos in the issue show the feature behind an environment variable. The variable name comes from the demo, not from any documentation, so expect it to change:

CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude

With the flag on, the demo shows a built-in /plugin-authoring skill that generates a plugin from a one-sentence description. The output lives in .claude/ as a plugin: a manifest, a hooks.json, and the .ts file with your hooks.

Verified against Claude Code 2.1.266

We did not take the architecture doc on faith. Two checks, both reproducible:

1. The engine is compiled in. The binary carries the event names, the error messages and the $ factory functions themselves. The surface is larger than the doc describes:

Beside those, the binary knows engine.create, plugin.register, turn.complete, and a classic.* namespace (classic.PreToolUse, classic.SessionStart, classic.SessionEnd, classic.Setup) — today's shell hooks, surfaced to function hooks as events. The engine.create withholding path even has its own diagnostics for a capability that was withheld but reached the host anyway, which is precisely the mechanism our admin-capability-lockdown hook assumes.

2. A catalog hook runs unmodified. We dropped security/secret-redactor's .ts into a plugin directory, changed nothing, and started Claude Code with the flag. A clean A/B: the same question with the flag off returned an unredacted .env; with the flag on, every one of the ten patterns fired. Redaction covered the Read tool as well as Bash, and a follow-up command containing a [REDACTED:…] placeholder was denied before execution, with the hook's own message. The register(on, options) and ($, e, next) signatures the doc describes are the ones the shipping binary calls.

The plugin layout

The architecture doc is precise about where things go. Hooks stay in hooks/hooks.json; one new key, modules, points at a file beside it. Your existing command hooks keep working next to the module.

my-plugin/
├── .claude-plugin/plugin.json
└── hooks/
    ├── hooks.json          # { "modules": ["./my-hooks.ts"] }
    └── my-hooks.ts         # export function register(on, options) { ... }

{
  "modules": ["./my-hooks.ts"]
}

The module may be .js, .ts, .jsx or .tsx. It exports a single register(on, options) function, where options is the plugin's userConfig. Because registration happens up front, claude plugin validate can list every event a plugin hooks before any hook runs.

Anatomy of a Hook

This is listing 1 from the architecture doc, the same "block rm -rf /" hook that the public docs use as the shell example, rewritten as a function:

export function register(on) {
  on("tool.call", ($, e, next) => {
    if (e.tool === "Bash" && e.command == "rm -rf /")
      return { deny: "Destructive command blocked by hook" }
    return next(e)
  })
}

Three parameters, three jobs:

• $ is the engine interface: everything a hook can see or do. It is an object of nouns, each an object of events: $.tool.call, $.ui.log, $.fs.read. It is the only door. The environment that runs your hook has no ambient filesystem or network, so what a plugin did is exactly the calls it made on $.
• e is the event: the argument the method was called with, as an immutable plain value. On tool.call it carries the tool name and its arguments as own fields. To change it, pass next a copy.
• next is the continuation. Calling it runs the next hook registered on the event and resolves to the result of the rest of the chain. You may call it once, many times, or never. It also carries next.event, next.origin (which plugin raised the dispatch), next.signal (an AbortSignal for the dispatch) and next.is(type, e) for narrowing under *.

An optional matcher between the event name and the callback narrows both the calls you see and the type of e. It is a partial of e, matched structurally; an array matches when any element matches:

on("tool.call", { tool: "Bash" }, ($, e, next) => { /* e.command is typed */ })
on("tool.call", { tool: ["Edit", "Write", "MultiEdit"] }, ($, e, next) => { /* any of the three */ })
on("ui.render", { component: "ToolUse", surface: "desktop" }, ($, e, next) => { /* one component, one surface */ })

Five placements, one event

Where a shell hook needs a pre-event and a post-event, a function hook decides where its logic runs relative to the real action by how it uses next:

Order Is Nesting

This is the part of the proposal that took the community a few replays to digest, and it is the part that matters most for security. Hooks registered on one event fold like middleware: on(X, A), on(X, B), on(X, C) becomes X = A(B(C(core))). The first plugin registered sits on the outside, sees every event first and every result last, and nothing beneath it can bypass it.

Organizations use exactly this. Managed settings list the plugins an administrator prepends (control) and appends (defaults). A prepended plugin can:

• Decide which plugins may exist at all, by hooking plugin.register.
• Decide which nouns exist on $, by hooking engine.create and returning the table without, say, http and process. A plugin below cannot call what is not there.
• See everything, by hooking *. That hook runs on every event, including every other plugin's own calls on $, so an audit log is one function.

// Listing 5 from the architecture doc: an audit log as one prepended hook.
on("*", ($, e, next) => {
  $.ui.log(`${next.origin} called ${next.event} at ${Date.now()}`)
  return next(e)
})

Three Hooks That Shell Hooks Cannot Express

The three snippets below are trimmed from components in the aitmpl.com catalog (linked in the table further down). They cover the three things the shell contract cannot do: rewrite, short-circuit, and draw.

1. Rewrite the input: npm to pnpm

on("tool.call", { tool: "Bash" }, ($, e, next) => {
  const rewritten = e.command
    .replace(/\bnpm\s+(install|i|add)\b/g, "pnpm add")
    .replace(/\bnpx\s+/g, "pnpm dlx ")
  if (rewritten === e.command) return next(e)
  $.ui.log(`[npm-to-pnpm] ${e.command} -> ${rewritten}`)
  return next({ ...e, command: rewritten })   // events are immutable: forward a copy
})

2. Short-circuit: cache WebFetch

const cache = new Map()

on("tool.call", { tool: "WebFetch" }, async ($, e, next) => {
  const key = `${e.url}\n${e.prompt}`
  const hit = cache.get(key)
  if (hit) return hit.result          // nothing below runs: no network call at all
  const result = await next(e)
  if (result && !result.deny) cache.set(key, { result })
  return result
})

3. Draw: a duration badge on every ToolUse

on("ui.render", { component: "ToolUse" }, async ($, e, next) => {
  const { Row, Badge } = $.ui.resolve(e)   // the surface's own elements
  const rendered = await next(e)           // whatever the engine drew
  return (
    <Row>
      {rendered}
      <Badge text={`${durationFor(e.props)} ms`} />
    </Row>
  )
})

The same JSX renders through Ink on the terminal and through the DOM on Desktop. A hook never receives rendered state, only props and the component to render, so the component names and props are declared public API in the doc, on the same footing as a tool's input schema.

10 Function Hooks in the Catalog

We wrote ten hooks against the API described in the architecture doc, one per pattern the proposal demonstrates. Each catalog entry is the real thing: the hooks.json with its modules key, and the TypeScript hooks-module it names, with a header comment that lists exactly which parts of $ are assumptions. Install any of them with the --function-hook flag, which needs claude-code-templates 1.29.5 or newer:

npx claude-code-templates@latest --function-hook security/block-destructive-commands

It writes the plugin layout above to .claude/skills/<name>/. Then start Claude Code with the flag. Pointing --plugin-dir at the directory is the path we tested and the one we recommend:

CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude --plugin-dir .claude/skills/block-destructive-commands

Claude Code should also pick the directory up on its own as <name>@skills-dir on the next session, once you trust the workspace — but that auto-load is the one part of this flow we have not confirmed on a real binary, so reach for --plugin-dir if the hook does not fire.

mkdir -p .claude/plugins/secret-redactor/{.claude-plugin,hooks}
cd .claude/plugins/secret-redactor
echo '{"name":"secret-redactor","version":"0.0.1"}' > .claude-plugin/plugin.json
echo '{"modules":["./secret-redactor.ts"]}' > hooks/hooks.json
# then drop secret-redactor.ts next to hooks.json and run:
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude --plugin-dir .claude/plugins/secret-redactor

Browse them all at aitmpl.com/function-hooks. The listing carries the same experimental banner as this article.

One more practical note: $.ui.log appears to be a TUI surface. In headless runs (claude -p, even with --debug) the hooks' log lines did not show up. That does not affect universal-audit-log, which writes through $.fs.append and deliberately skips ui.* events, but it does mean a hook that reports only through $.ui.log will be silent in CI.

What the Thread Is Still Asking

The issue collected serious feedback within a day, much of it from people who run dozens of hooks in production. If you are deciding whether to invest, these are the open questions worth tracking, none of which the doc answers yet:

• Fail-open or fail-closed? If a hook three deep throws, does the action proceed or get blocked? For a redaction hook, fail-open is worse than no hook at all. Several commenters want this declared per hook.
• Hang budget. What cancels a hook whose promise never settles? Partly answered by the binary: there is reentrancy detection, and it is specific. Call $.model from a prompt.submit hook and 2.1.266 tells you the call "would wait on the turn this hook is holding", then names the ways out — answer { text }, call next(e), or defer to a later event such as turn.complete. There are budgets too (a per-plugin model-token budget that can be spent). What is still unclear is the generic case: a plain await on something slow that is not the model.
• How far does $.fs reach? Many real hooks read ~/.secrets/ and write state outside the project. The methods exist in 2.1.266 — readFile, writeFile, listDir, exists, stat, ancestors — but what paths they will actually accept, and under whose permission settings, is not something you can read off a binary. The author's stated intent is not to restrict plugins but to route everything through $ so admins can audit, allowlist and deny.
• Do MCP tool calls go through tool.call? If they reach the model outside $, the biggest gap in today's guards stays open.
• Interleaving with command hooks. Nobody migrates 150 hooks in one release. Does a command hook's deny win over a function hook that called next?
• Testing. A documented fake $ and a fixture format would decide whether existing test suites port or get rewritten.
• Forward compatibility. One maintainer measured that an unknown key in hooks.json makes older Claude Code versions drop the whole file silently. A modules key needs to be skipped, not rejected, by builds that predate it.

Should You Care Now?

If you only need allow/deny, your shell hooks are fine and will keep working; the doc is explicit that command hooks run beside function hooks, not instead of them. Function hooks earn their complexity in three places: when you need to change what a tool receives or returns, when you need state or a UI, and when you are an administrator who needs a control that cannot be argued with. If any of those is you, the ten catalog hooks are a concrete starting point, and the issue is where your opinion changes the outcome.

Created by Daniel Ávila