Claude Code Mods: Rewrite Behavior, Build Custom UIs — Anthropic’s Version of dsh Is Here

Anthropic just dropped Claude Code Mods, a game-changing new feature that lets you modify how Claude Code works with just a few lines of TypeScript — even build fully custom UI elements directly into the interface.

The official definition says it best: “Mods let you rewrite or replace Claude Code’s behavior, and even render custom UI. Under the hood, mods are hooks that are distributed alongside plugins. Each mod is a small JavaScript or TypeScript module that runs inside your session and can observe every event that occurs.”

Put simply, mods can intercept, modify, or block literally every action: tool calls, permission requests, even UI rendering passes. Most notably, Claude Code’s own built-in /diff tooling and AGENTS.md support are already implemented as mods. That means you can disable them entirely, or swap them out for your own custom alternatives.

Three Examples To See What Mods Can Do

The official Claude Code blog shares three working example mods, with full code available in the anthropics/claude-code repository:

1. Token Weather

Adds a status bar directly above your input box that shows real-time context window usage. It displays a sunny ☀ icon when you’re under 25% usage, and a lightning ↯ icon when you’re over 90% and compression is imminent. The entire thing clocks in at just 80 lines of code.

2. Blast Radius

Intercepts dangerous commands like rm -rf or git reset --hard before Claude runs them, calculates exactly which files will be affected, and opens a side panel with "Continue" and "Cancel" buttons to let you approve or block the action (1 to proceed, 2 to cancel).

3. Replay Theater

Records every file edit Claude makes through your session, and lets you open a player-like panel after a chat turn to step back through every diff change one by one, just like rewatching a video.

All mods are built on the same foundational hook system: every time Claude Code takes an action, it emits an event that mods can observe, modify, or reject entirely. Mod state is stored in $.state, which persists across hot reloads so you can tweak your code without restarting your entire session. Hot reloading works out of the box, too — just edit your code and Claude Code reloads it automatically. Mods are enabled by default in Claude Code 2.1.287 and newer.

If you don’t want to write a mod from scratch, you don’t have to: Claude can write it for you. The official workflow lets you describe the mod you want in plain natural language, and Claude will generate the code and hot load it directly into your current session.

A Counterintuitive Design Choice

We’re used to configuring our tools, but mods take it a step further: they let you rewrite the tool itself. Anthropic even moved all their own built-in features to mods intentionally, so you can adapt the tool to fit how you work instead of the other way around. This design assumes users have both the ability and desire to customize their tools — which means more freedom for individual developers, but new oversight challenges for teams.

As one developer put it: “Any mod that can rewrite your tool means every plugin lives inside your trust boundary. It’s awesome for tinkerers, but a headache for teams that need to audit every plugin they use.”

Mods run with the same full machine access that Claude Code has, so the official docs explicitly warn users to only install mods from trusted sources. For team and enterprise plans, there are built-in security mods that block user-installed mods from overriding existing permission rejection rules.

The Community’s First Reactions

Right after the announcement, developers started sharing their first custom mod creations. Some folks have already built new functionality that fits their unique workflows.

Others shrugged it off, joking that they’d already customized their Claude Code setups beyond recognition. Many have also been asking why Claude subscriptions can’t be used across other AI tools besides Claude Code. It’s clear that developers have wanted more open, customizable AI clients for a long time.

What This Means For The Future Of AI Coding Tools

In the official announcement, Anthropic says Claude Code should “feel like it’s yours.” Mods are the first step toward that goal. It pushes customization from simple config edits all the way to full source-level changes to the client itself. You don’t have to wait for Anthropic to ship a new feature — you can just build it yourself.

This will feel familiar to anyone who’s used deepseek harness: instead of getting a static, one-size-fits-all tool, you get an evolvable, fully customizable harness framework. When every developer can rewrite their client, the value of the tool itself shifts toward the underlying model and the broader mod ecosystem. Claude Code as an entry point is becoming a blank canvas — and developers get to decide what goes on that canvas.

If you want to build your own mod, check out the official getting started guide here: Getting started with Claude Code mods. Or just ask Claude to write it for you — it supports generating full working mods from natural language descriptions.