Anthropic adds Claude Code Mods for JavaScript and TypeScript plugins

• Claude Code Mods are JavaScript or TypeScript plugins that can replace agent behavior and draw custom UI, as described in the launch demo.
• In-process hooks expose conversation context, subagent spawning, and structured results, according to the interview clip.
• Claude Code can build a mod from a prompt and hot-reload it into the session, following the prompt-to-mod walkthrough.
• Community tooling already includes a free mod-building skill that daniel_mac8 shared during the early preview.

Anthropic's roughly 80-line Token Weather example turns context usage into a weather forecast above the prompt. The published built-in source reveals that AGENTS.md support and the /diff pane use the same extension mechanism.

Mods are enabled by default in Claude Code 2.1.287 or later, according to Anthropic's October 1 guide. A September 25 independent breakdown still documented the preview flag CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1.

The API can change between releases. Every mod load writes declarations into .claude-plugin/types/, which the guide identifies as the authoritative API for that build.

A mod uses the standard plugin package layout specified in the guide:

• .claude-plugin/plugin.json: the plugin manifest.
• hooks/hooks.json: a modules array naming exactly one module.
• The module: exports register(on, options) and attaches handlers with on(event, matcher?, hook).

Each handler receives ($, e, next): the host API, plain event data, and a continuation. Hooks form a middleware chain, with Claude Code's normal behavior at the bottom.

A handler has three moves:

1. Observe: await next(e), inspect the result, then return it.
2. Rewrite: pass a modified event to next.
3. Answer: return directly without calling next, denying an action or supplying the implementation itself.

That third move lets a plugin serve a tool or slash command instead of passing it through to the core.

Thariq Shihipar, on Anthropic's Claude Code team, outlined three in-process capabilities in the podcast conversation:

• Inspect conversation turn counts and token usage.
• Spawn subagents from the plugin.
• Parse their results and return structured output.

For usage widgets, $.session.usage() exposes context.tokens, context.window, and context.percent, as documented in the Token Weather tutorial. The token figure represents the input the last response was answered over; the call only sends a token-count request when a breakdown is requested.

The biggest unlock is component-level UI interception. A ui.render hook can replace the AbovePrompt band or draw a Pane, using constructors returned by $.ui.resolve(e) in the official example.

The constructors vary by surface, and JSX is supported with h as the factory. The guide demonstrates two different uses:

• Token Weather: context percentage, tokens used versus window size, a 12-turn sparkline, and the last turn's token change.
• Blast Radius: holds a risky Bash call, measures what it would affect, and opens a Proceed/Cancel pane. Its rm -rf build demo previews nine files totaling 1.1 MB before execution.

Claude Code can write a weather-forecast mod from a plain-language request, then display it when the turn ends after hot reload is approved, according to the walkthrough.

Reloading re-runs register and fires session.start again. Anthropic's working example spells out the lifecycle details:

• Module-level variables reset on reload.
• Named values in $.state live in the host for the session and survive reloads.
• Calling $.state.get during rendering subscribes that drawing; a later $.state.set triggers a redraw without $.ui.invalidate.
• Prompt-generated mods load only in the current session, and their temporary folders are cleaned up later. Keeping one requires copying the folder out and installing it as a plugin.

Development loading also supports claude --plugin-dir ./token-weather, with saves reloading the watched module in place.

A free claude-mod-builder skill appeared during the early preview, packaging the function-hook model for Claude to use.

Anthropic's tutorial provides two separate checks:

• claude plugin validate ./token-weather: reports hooked events, API calls, and state reads and writes. State keys must be declared in the plugin's PluginState type contract, referenced by the manifest.
• claude plugin test ./token-weather: runs *.test.ts files against the real Claude Code runtime. Test hooks sit beneath the mod in the chain, where they can stub usage figures and other host responses.

The built-in Mods README adds a useful testing detail: a call left unanswered by those test hooks throws an error naming its event.

A GitHub repository with .claude-plugin/marketplace.json supplies an ordinary plugin marketplace. The documented installation sequence is:

Anthropic describes directory submissions as forthcoming, available once the directory accepts plugins containing Mods. That differs from the demo post, which describes Claude Directory sharing as already possible.

A mod's sandbox has no DOM or Node; filesystem, HTTP, process, and other external operations go through $. Anthropic nevertheless warns in its guide that installed mods run on the user's machine with the same access Claude Code has.

The public function-hooks design thread demonstrates how authority composes:

• The first registered plugin wraps those registered afterward.
• An administrator can remove capabilities from $, preventing plugins beneath that layer from invoking them.
• A wildcard hook on * observes every event, including plugins' own API calls, enabling an audit log.

Claude Code's published Mods source lists four built-in plugins, with their tests:

• agents-md: loads AGENTS.md when a project has no CLAUDE.md by default. Options also support loading both files, using only CLAUDE.md, or retaining only managed instructions.
• diff: renders uncommitted changes beside the transcript, file by file, and refreshes as Claude edits files and runs commands.
• telemetry: supplies first-party analytics hooks to built-in plugins, refuses installed plugins, and sends nothing when Claude Code analytics are off.
• sec-default: protects organizational hooks, prompt content, managed settings, tool policy, and deny rules from user-installed plugins. It sits outermost on managed machines or for Team and Enterprise organizations unless managed configuration changes that placement, and adds no policy of its own.