Anthropic dropped custom themes for Claude Code on November 13 — and they’re built on the mod system it launched six weeks ago that most developers haven’t touched yet. If you’re still running Claude Code out of the box, you’re leaving a lot on the table. Mods let you intercept prompts, block risky commands, redact secrets from tool output, and now skin the entire interface. Here’s what they actually do and what’s worth building.
Mods Are Middleware, Not Extensions
The framing matters. Claude Code Mods are not add-ons that sit alongside the agent — they run inside its execution pipeline. Think Express middleware for AI agents: small TypeScript or JavaScript handlers that hook into events the agent already fires before a prompt reaches the model, before a tool call runs, or when the UI redraws.
They shipped in v2.1.287 (October 1) and are active by default. Install them with /plugin install name@marketplace in a session or claude plugin install from your shell. The official marketplace sits at 317 plugins as of November 9. The unofficial ecosystem grew to 10 of 15 GitHub Trending spots within 48 hours of launch.
Four Things You Can Actually Do
- Rewrite prompts before they hit the model. The
onPromptSubmit()hook intercepts the raw prompt. A simple mod can detect unstructured brain dumps and reformat them into Goal / Constraints / Acceptance Criteria before Claude ever sees them. - Hold or redirect tool calls. A mod can pause any tool call, show you what it would do, and require explicit confirmation — or reroute it to a different model entirely.
- Redact secrets from tool output. Mods strip environment variables and API keys from tool results before Claude processes them — useful when tools return config files you don’t want sitting in context.
- Add custom UI. Panels, buttons, CI status feeds, tabs. Within 48 hours of launch, someone had Tetris running inside the terminal.
Start with Anthropic’s Own Examples
Four official mods from Anthropic are worth installing before building anything custom.
Blast Radius is the most practically useful. It intercepts commands with destructive potential — rm -rf, force pushes, database migrations — and shows exactly what they’d touch before asking you to proceed or cancel. It’s the guardrail that probably should have shipped with Claude Code on day one.
Token Weather draws a context-window forecast above the prompt input. Clear skies when you have headroom, showers as context fills, a full storm when you’re close to the limit. Around 80 lines of TypeScript. Surprisingly effective for avoiding context surprises mid-task.
Replay Theater records every file edit made during a turn, then lets you step through the diffs one at a time after the turn ends. Essential for multi-file change reviews.
You Should Know (built-in since October 3) deploys a monitoring agent that surfaces important things you might skim past in Claude’s output. It’s Anthropic eating its own cooking — this feature is itself a mod, not core code.
Themes: What Changed on November 13
Themes are mods that call $.theme.set(). Colors require v2.1.324+; animated backdrops require v2.1.319+. Run claude --version before attempting either. One theme is active per session — any mod that calls the theme API replaces what’s currently running.
The recommended workflow: ask Claude to generate the theme for you. Describe what you want, it writes the TypeScript, you install it. Currently desktop-only; cloud session support is on the roadmap.
The Security Reality
This is where most coverage is too polite. Mods are not sandboxed. They run inside Claude Code’s process with your full machine permissions. That means a mod can read and write any file your account can access, pull environment variables and API keys, approve tool calls without prompting you, and consume your API quota silently.
This isn’t theoretical. On October 29, trojanized builds of the official Claude Code VS Code extension appeared on Cursor’s marketplace. Roughly 5,000 users were exposed over 11 hours. A separate security audit by NeuralTrust found 26.1% of marketplace plugins contained at least one vulnerability; 13.3% attempted data exfiltration.
The mitigation stack: run claude plugin validate before installing anything, use --safe-mode to launch without mods when you need a clean environment, and treat mod installs the same way you treat installing any executable from the internet — because that’s exactly what they are. Organizations can configure admin-level allow-lists to restrict which mods load.
Getting Started
- Confirm you’re on v2.1.287 or later:
claude --version - Browse Anthropic’s samples in the claude-code-playground repository. Install one and see how it hooks in before writing anything.
- Write your first mod by asking Claude to write it. Describe the behavior you want — rewrite noisy prompts, block specific file paths, add a word count badge to the UI — and have it produce the TypeScript. The official mods documentation covers the full hook API.
The Bigger Picture
The mods system is a platform play, not a feature drop. Four of Claude Code’s own built-in capabilities — diff, agents-md, sec-default, telemetry — are implemented as mods. Anthropic is using the same API it shipped to developers. VS Code became dominant partly because its extension ecosystem made every workflow possible. Anthropic is making the same bet on Claude Code.
Whether it pays off depends on whether the ecosystem matures responsibly. A 26% vulnerability rate across 317 plugins is not a healthy baseline. That’s a solvable problem — the tooling exists (claude plugin validate, pinned commit SHAs, org allow-lists) — but it needs solving before the marketplace scales to thousands of plugins and the attack surface grows with it.

