Claude Code Mods give plugins interventions and interfaces

What this is about

Anthropic is extending Claude Code with what it calls Mods. These are plugins with a JavaScript hooks module. They can observe events inside a Claude Code session, influence tool calls, and draw their own controls directly in the terminal. A mod can therefore intervene much more deeply than a simple bundle of instructions.

On October 1, 2026, Anthropic added three examples to its public Claude Code Playground: Blast Radius, Replay Theater, and Token Weather. They require Claude Code 2.1.287 or later. Anthropic explicitly describes the Playground as a collection of examples rather than an officially supported product. The source code is available under Apache 2.0.

What Claude Code Mods actually do

A regular Claude Code plugin can package skills, subagents, conventional hooks, or MCP servers. A mod adds a hooks module made of JavaScript functions. That module can react when a session starts, a tool is called, or a turn completes. It can also render a side pane or a line above the input prompt.

The three examples demonstrate different levels of intervention. Blast Radius first holds risky shell commands such as rm -rf, git reset --hard, force pushes, or database migrations. Before execution, it works out which files, commits, or migrations would be affected. The human then chooses between Proceed and Cancel through buttons.

Replay Theater observes file changes and presents them as a sequence after a turn. Users can move forward and backward through the individual diffs. The mod does not alter the work; it makes the result easier to review. Token Weather reads real context-window usage values and displays them as a weather state, a percentage, and a history of the last twelve turns. It does not call a language model to do this.

Why it matters

Many Claude Code extensions have so far been mostly invisible logic: instructions, commands, or background processes. Mods combine that logic with visible interaction in the terminal. This matters when a coding agent works for a long time, changes many files, or runs commands whose effects are difficult to undo.

Blast Radius illustrates the main difference. A hook can do more than log an event after it happens; it can hold a tool call and wait for a human decision. Replay Theater turns a long list of edits into a short visual review. Token Weather makes it clear when a session has consumed a large share of its context and automatic compaction is getting closer.

That power also raises the trust stakes. According to Anthropic, plugin code runs with the privileges of the signed-in user. A third-party mod can therefore read files, launch processes, or access the network if its code is designed to do so. A marketplace entry is a catalog listing, not a security seal. Before installation, users need to inspect the manifest, hooks, dependencies, and network behavior.

In plain language

Think of Claude Code as a workshop. A normal plugin puts extra tools in the cabinet. A mod can also fit a guard over the table saw, mount a camera that replays the last steps, or add a fuel gauge. That can make the workshop clearer and safer, but only if you trust the accessory you installed.

A practical example

A developer asks Claude Code to move an application to a new database version. The agent changes twelve files in one turn and then wants to run a Prisma migration. Blast Radius pauses the command and displays the pending migration. The developer cancels for now, opens the twelve edits in Replay Theater, and checks the five most important diffs.

Token Weather meanwhile reports 78 percent of a 200,000-token context window. Instead of adding three more tasks, the developer starts a fresh session after the review. She then allows the migration in a test environment. This scenario is fictional, but it uses exactly the behavior documented for Anthropic's three open mods.

Scope and limits

First, Blast Radius is only a safety net. It checks selected patterns in a Bash call, but it does not parse the full shell language. Hidden calls through eval, bash -c, aliases, scripts, or tools such as xargs can bypass detection. Once Proceed is selected, the entire command runs unchanged. Hard blocks still belong in permission rules or a sandbox.

Second, mods share limited terminal space. Multiple extensions can compete for the same band above the prompt. Replay Theater also keeps its history only in the current session's memory, while Token Weather updates only after a complete turn.

Third, the three examples are not supported Anthropic products. Their own documentation names untested paths and specific limitations. Teams should validate mods with claude plugin validate, review their source, and test them in a disposable project first. For mods from third-party marketplaces, a security review matters more than a convenient one-click install.

SEO & GEO keywords

Claude Code Mods, Anthropic, Claude Code Plugins, JavaScript Hooks, Blast Radius, Replay Theater, Token Weather, Coding Agents, Terminal UI, Plugin Security, Context Window, Human in the Loop