Image: generated by METAL AI
Summary
- On September 3, X account bcherny, associated with Claude Code, previewed early plans to make Claude Code far more extensible than it is now.
- The post pointed to GitHub issue #91870 for details, and that issue reportedly contains a "Function Hooks" proposal.
- Claude Code has already built out extensibility through skills, plugins, hooks, and MCP servers, and this teaser suggests an even bigger shift is coming.
On September 3, X account bcherny, which is associated with Claude Code, shared early thinking on an idea that would make Claude Code far more extensible than it is today — and put the question directly to developers: would you use this? The teaser stands out because Claude Code, the AI coding agent built by Anthropic, has already been stacking up extensibility through skills, plugins, and hooks.
To put it plainly: Claude Code is an AI coding tool that already lets you bolt on outside functionality through skills, plugins, hooks, and MCP servers. This latest teaser signals that the extension model itself is about to change in a bigger way than what's come before.
Claude Code writes and fixes code on a developer's behalf, working inside a terminal or development environment. It's less a chatbot that spits out answers and more an AI agent that creates files and runs commands, and over the past several months it has steadily broadened how outside functionality gets attached to it — through skills, plugins, and hooks.
The whole message fits in a tweet
Bcherny's post isn't long. It opens with "Your input needed — would you use this?" and describes early thinking on making Claude Code far more extensible. The post goes on to call the idea "a bit unconventional, and pretty exciting." The tweet itself doesn't spell out exactly how the extensibility would work; instead it points to github.com/anthropics/claude-code/issues/91870 for the details.
The link doesn't show any content
Following that link, though, doesn't load the issue itself — it lands on the repository's issue search page with a message saying no results were found. That appears to be because a trailing ellipsis got tacked onto the end of the URL in the tweet. As a result, we haven't been able to confirm what extensibility approach issue #91870 actually describes.
How Claude Code has already been expanding
To make sense of this teaser, it helps to look at how Claude Code has been growing its extensibility so far. The /design skill, released as a research preview last month, brought an artboard-style workflow into the coding tool, letting it generate multiple UI drafts. The plugin feature, unveiled earlier this month, bundles skills, subagents, hooks, MCP servers, and some configuration values together so a whole team can deploy them at once. Hooks, introduced around the same time, force a defined piece of code to run every time a tool is called, making it impossible to skip certain rules.
| Extension type | What it does | Rollout |
|---|---|---|
| Skill | Teaches a specific workflow to automate repetitive tasks | (/design research preview) |
| Plugin | Bundles skills, subagents, hooks, and MCP servers for deployment | 2026-09-03 (Claude Academy, lesson 9) |
| Hook | Forces defined code to run at each tool call | 2026-09-03 (Claude Academy, lesson 5) |
If you want to leave feedback directly, the address given in the tweet is github.com/anthropics/claude-code/issues/91870. The version we accessed didn't display the issue content, but logging in first might produce a different result.
Editor's take
The phrase "way more extensible" landing so soon after skills, plugins, hooks, and MCP were all rolled out isn't something to brush off. Up to now, extending Claude Code has mostly meant layering procedures that developers already wrote onto Claude. Skills carried the procedure, plugins bundled those procedures together, and hooks enforced when things had to run. Taking things a step further usually points in one direction: opening up a structure where a third-party ecosystem outside Claude Code can plug directly into it — something closer to a browser's extension store.
Anyone who's watched this play out before will find the pattern familiar. Slack opened up webhooks and bot APIs before it launched an app directory. VS Code refined its command palette and APIs before opening an extension marketplace. Claude Code looks like it's following the same sequence — laying the basic plumbing with skills, plugins, and hooks first, and now moving on to the question of who gets to build on top of that plumbing, and how.
Teams already running Claude Code don't need to change anything right now. But if your organization has built its own skills or plugins, it's worth designing them with an eye toward the possibility that, once this idea takes shape, sharing them externally — or pulling in someone else's — could get a lot easier than it is today. Right now the GitHub issue page is inaccessible, so the details remain unclear, but there's a good chance Anthropic will turn this into a more concrete announcement or research preview within the next few weeks.
Comments