3 Claude Code Mods That Change How the Agent Behaves
You are not only adding features to Claude Code anymore. With mods, you can change how Claude Code itself behaves — tool calls, prompts, turns, and the terminal UI (Anthropic announcement).
Anthropic's getting-started guide (published Oct 1, 2026) does not pick three random demos. It ships three sample mods that each solve a different job: context awareness, risky-command control, and post-turn review (Getting started with Claude Code mods). This post walks through those three — Token Weather, Blast Radius, and Replay Theater — and what each one is for.
Key Takeaways
- Mods are TypeScript/JavaScript plugins that hook Claude Code events and can draw their own UI. They require Claude Code v2.1.287+ and are on by default (mods guide).
- Token Weather shows a live "forecast" of how full the context window is above the prompt — Clear → Storm → Compact soon.
- Blast Radius holds risky Bash (
rm -rf, force push,git reset --hard, migrations, …), previews impact, then Proceed / Cancel. - Replay Theater records Edit/Write in a turn and lets you step through diffs with
/replayorr. - Mods are not MCP tools or simple slash commands: they can observe, rewrite, or answer events in the hook chain.
What Claude Code Mods Actually Are
Answer first: a mod is a Claude Code plugin whose behavior lives in a JS/TS module. hooks/hooks.json points at one module; that module exports register(on, options) and attaches hooks to events such as tool.call, turn.complete, ui.render, and command.run.
Hooks form a middleware chain. Your hook runs, next(e) passes the event onward, and Claude Code does its default work at the end. A hook can:
| Move | What it does | Example |
|---|---|---|
| Observe | Call next, then look at the result | Record every file edit; read usage after each turn |
| Rewrite | Pass a changed event into next | Soften or redirect a command |
| Answer | Return { deny: "…" } without next | Refuse a Bash call; serve a slash command yourself |
That is the real difference from MCP servers (extra tools for the model) or statusline scripts (display-only). Mods sit inside the session, keep state, open panes, and can block or rewrite what would have happened.
Anthropic also notes that some built-in Claude Code features — including AGENTS.md support and the /diff pane — are themselves implemented as mods, with source under mods/ in the public anthropics/claude-code repository (mods guide).
Prerequisites
claude --version # need 2.1.287 or later
Mods are enabled by default. The API can still change between releases; when Claude Code loads a mod it writes type declarations into the mod's .claude-plugin/types/ folder — those types match your build (mods guide).
Treat third-party mods like any package that runs on your machine: read the repo, install only from sources you trust.
The Three Sample Mods at a Glance
| Mod | Problem it solves | Demo impact |
|---|---|---|
| Token Weather | Context window tracking | Strong for long sessions |
| Blast Radius | Safer tool execution | Highest "wow" for demos |
| Replay Theater | Review what Claude just changed | Great for learning and video |
1. Token Weather — See Context Before It Surprises You
Value: you no longer have to guess how full the context window is.
Token Weather reads $.session.usage() after each turn and draws one line above the prompt: weather icon + word, percent used, tokens out of the window, a short sparkline of recent turns, and how much the last turn added (mods guide, mods showcase).
| Used | Forecast |
|---|---|
| under 25% | Clear |
| 25–49% | Cloudy |
| 50–74% | Showers |
| 75–89% | Storm |
| 90%+ | Compact soon |
Conceptual readout:
Context Window
██████████████░░░░░░ 72%
Used: 72%
Remaining: 28%
In a real session the band is a single status line (for example ☂ Showers 67% of context 134.4k / 200k plus a sparkline). On large repos or multi-hour coding sessions, that line is the difference between "Claude got confused" and "we were already at Storm."
The reference implementation is roughly 80 lines. Anthropic's guide also shows a shortcut: paste a natural-language prompt into Claude Code and let it scaffold token-weather with hot reload in the same session (mods guide).
2. Blast Radius — Intercept Dangerous Shell Before It Runs
Value: instead of giving Claude blanket trust, add a guard that interrupts critical shell.
When Claude calls Bash with patterns such as rm -rf, git reset --hard, git clean, a force push, or a database migration, Blast Radius holds the call, estimates what would change (via dry-run style checks like git clean -n, git status --porcelain), and opens a pane with Proceed and Cancel (mods guide).
Imagine Claude wanting:
rm -rf node_modules
or:
git push --force
What you get is closer to:
⚠ Risky command detected
git push --force
Potential impact:
• Remote branch will be rewritten
• Existing commits may be lost
[ Cancel ] [ Proceed ]
In Anthropic's recording, Cancel returns a refusal to Claude with a reason; Proceed runs the original command. Hotkeys: 1 = Proceed, 2 = Cancel.
Important honesty from the official guide: Blast Radius is a safety net, not a permission system. It classifies command text, so aliases, $(…), or scripts that call rm can slip past. Use Claude Code permission rules when you need a hard block (mods guide).
Under the hood this is the answer move on tool.call for Bash: measure impact, wait for a UI decision, then either next(e) or return { deny: "…" }.
3. Replay Theater — Step Through Claude's Last Edits
Value: "What did Claude just do?" becomes a step-through, not one giant diff.
Replay Theater records every Edit and Write in a turn (file path + before/after text). When the turn ends, a hint appears above the prompt. Press r or run /replay to walk edits one diff at a time — numbered steps, Prev / Next / Close (mods guide, mods showcase).
If Claude touched eight files in one turn:
/replay
Step 1 → app/api/auth.ts
Step 2 → lib/auth.ts
Step 3 → middleware.ts
Step 4 → app/login/page.tsx
…
[ ← Previous ] [ Next → ]
It never blocks or rewrites edits — pure observe. That makes it ideal for demos, teaching, and learning in what order the agent solved a problem.
How the Three Fit Together
- Token Weather answers: How much context is left?
- Blast Radius answers: Should this shell command run?
- Replay Theater answers: What files changed, and in what order?
Together they show the mods thesis: you are not only bolting on MCP tools — you are reshaping Claude Code's loop and UI. If your stack already leans on MCP for browser, docs, and research, keep that layer (5 Claude Code MCPs); mods sit underneath and change how the agent acts, not only what it can call.
Try or Build a Mod Yourself
Local plugin dir (fastest for learning)
claude --plugin-dir ./token-weather
Edits hot-reload while the session stays open.
Share via marketplace (three commands)
Mods install like any other plugin. From Anthropic's guide:
/plugin marketplace add your-org/my-mods
/plugin install token-weather@my-mods
/reload-plugins
Or from the CLI with a local marketplace folder:
claude plugin marketplace add ./my-mods
claude plugin install token-weather@my-mods --scope user
Validate and test with:
claude plugin validate ./token-weather
claude plugin test ./token-weather
Shortcut: describe the mod to Claude
You do not have to memorize the API first. Anthropic's Token Weather shortcut is a paste-ready prompt that asks Claude to create the mod and enable hot reload. Change the "what it should show" lines and it becomes your mod (mods guide).
FAQ
Do mods replace MCP?
No. MCP adds tools the model can call. Mods hook the Claude Code runtime (tools, turns, UI). Use both: MCP for capabilities, mods for behavior and guardrails.
Are sample mods built into every install?
Anthropic documents Token Weather, Blast Radius, and Replay Theater as the teaching samples on claude.dev/mods and in the getting-started guide. Built-in product features that are also mods (for example /diff, AGENTS.md) live under mods/ in anthropics/claude-code. Treat community marketplaces separately and verify before install.
Is Blast Radius enough to run Claude unsupervised?
No. Pair it with permission rules and human review for production or destructive work. The official guide is explicit that classification by command string is incomplete.
Which Claude Code version do I need?
2.1.287 or later (mods guide).
What to Do Next
- Upgrade to Claude Code ≥ 2.1.287.
- Run Token Weather first — lowest risk, immediate feedback in long sessions.
- Add Blast Radius before you let Claude loose on cleanup or git history.
- Use
/replayafter multi-file turns until you trust the agent's edit order. - When you have a question the samples do not answer ("cost meter?", "block
terraform apply?"), describe that mod to Claude and keep the folder as a plugin.
Mods turn Claude Code from a fixed agent into something you can reshape. Start with Anthropic's three samples, then build the fourth for your workflow.
Related articles
- ElevenLabs MCP Claude Code: Hosted Install, OAuth, and Creative ToolsInstall the hosted ElevenLabs MCP in Claude Code with one HTTP command, complete OAuth via /mcp, then generate speech, music, images, and video—or manage voice agents—without a local server or API key.
- Find LinkedIn Jobs with Claude + Apify (Free Credits)Connect Apify in Claude Connectors, paste one prompt with your resume, and get a ranked LinkedIn jobs table—fit score, salary, date, apply link, plus resume tweaks per listing.
- OneSignal Expo: Production Push Setup and Common ErrorsShip OneSignal push on Expo—development builds, permissions, tokens, Expo Go limits, APNs/FCM, External IDs, and the errors that waste hours, plus soft kit wiring notes