We're exploring a new way to let you extend and customize Claude Code: Function Hooks. Here's a couple videos showing what you'd be able to do. It hasn't shipped yet, we'd love feedback on this on our GitHub issue.

Can't you just add support for AGENTS.md?

Nice functionality, awful video - don't speak claudish. It's bad enough in writing, hearing it from a person is cringe inducing.

I can't help but notice the video's script, it's cleary made by AI, which is okay, but it's the most generic AI script ever, it's very distracting.

Not surprised to hear Claudish spoken in the video. Please fix Opus 5. It’s hurting everyone, even your video voiceovers.

Seems cool. I wish you didn’t subject me to the obnoxious Claude writing for the video script though.

"we'd love feedback" as the most polite way to say "please find our bugs before we ship this"

How about instead you give our usage back 😆

How about hooking the pricing page to give our usage back

More control over context compacting/editing would be great. Stratified context (system, highly persistent, volatile memory, literal history) that typically compacted ancient literal to volatile would mean less savings on compact but plausibly maintain state better

@grok can I have this in English please?

Extensions are great until they turn a lightweight CLI tool into a bloated IDE. You're trading speed for complexity; let's see if you can keep the latency under 100ms once users start chaining these hooks.

This is why I quit using Claude Code... Just saying hello and you shoe horn 35k+ tokens into the context window instantly... Unless this is constant hot cache its a huge hit to usage and cost to pay for 30k+ tokens each request... This is why Pi is the goat and Codex.

anything but raising the limits

wow Function, what an ancient word.

Oh god, you talk just like Claude. Poor anthropic employees, now speaking claudese

Can we customize the ability to copy text in the CLI to not have line breaks! For the love of God it needs to be fixed

I SUBSCRIBD TO X FOR 20 AND DOWNLOADED VIDEOS . IM ABOUT TO KEEP LEARNING AND APPLYING UNTIL I LEARN WHAT I MUST KNOW TO SURPASS MY LIMITS

Feedback from having built this by hand: what I needed most was a hook that can halt and hand control back, not just observe or transform. I run an AI reviewer over every PR under a policy that a rejected finding drops the executor back to per-operation approval. 171 written verdicts so far — 158 accepted, 11 rejected, 2 deferred. Each of those 11 stopped the chain and waited on a human. Implementing that meant reconstructing state the harness already had. A hook that could return "stop, ask the human, resume here" would have replaced the whole thing. Second ask: whatever fires needs to leave an artifact. My halts write a rationale inline into the source — 34 of them, greppable months after the session is gone. A hook that changes behavior and leaves no trace is a debugging problem you meet at volume, not on day one.

You should be researching more ways to be open to your community, allow third party harnesses & connect with Claude.

Is it possible for these plugins to be pulled into the path so you can test them out or use them in your shell directly?

The most useful hook isn’t just another place to run a script; it needs a clear lifecycle contract. Give it structured context, explicit permission boundaries, and predictable failure/retry behavior so teams can extend workflows without turning every hook into invisible global state.

One important detail to define before shipping: what happens when a hook throws, hangs, or mutates its input, and whether administrator policy can reliably take precedence over user plugins. Predictable failure and ordering rules will matter for managed deployments.

Change what Claude shows, curious how

Function Hooks still pre-ship and they're asking for feedback first. rare

Wont the high entropy thing mess with log files that uses hashses and uuids? @grok

NOONE cares, when you can't even utilize the product for more than a few hours....what is the point

Do you really still need to restart to pick up these things? Holy shit.

even the devs aren't using opus 5

being able to intercept and format tool calls before execution solves a huge headache for running strict local linting and security policies on generated code

Hooks that redact secrets or change what the model sees aren't just plugins - they're a containment layer most agent setups skip entirely. That's the real permission boundary, not a system prompt. #AgenticAI #LLMOps @AnthropicAI

We actually went deeper on how these hooks allow you to customize Claude Code through TypeScript: