Claude Code Mods (RoboNuggets)
🗒️ Description
A 13-minute YouTube walkthrough by Jay of RoboNuggets on mods, the newest extension surface in Claude Code. A mod is a small add-on that runs inside Claude Code and changes how it looks and behaves: extra buttons, side panels, a progress bar, even a mini-game. You describe the feature you wish Claude Code had and it builds that feature into itself.
🚀 TL;DR
- Skills teach Claude how to do a job. Connectors (MCP) plug it into another app. A mod changes Claude Code itself, including what is drawn on screen.
- Nine mods are worth a look: a subagent progress panel, themes, a “you should know” watcher, a live file tree, a prompt-cache warning, a destructive-command gate, a feedback-to-rule catcher, an in-terminal browser and an edit replay.
- Making your own is “just ask Claude”: a built-in skill knows how mods are put together.
- Source caveat: the channel sells a paid community and a PDF of the nine prompts, so read it as an enthusiastic walkthrough, not a review. Version numbers and dates below are as the video states them.
đź”— Links
🧩 What a mod is
| Does | Runs | |
|---|---|---|
| Skill | Tells Claude how to do a job (a handbook) | As instructions Claude reads |
| Connector (MCP) | Plugs Claude into another application | Outside Claude Code |
| Mod | Changes Claude Code’s own UI and behaviour | Inside Claude Code |
How it works. Every time Claude Code does something (runs a command, starts a helper) it emits an event. A mod is a small piece of code that listens for events and can do one of three things with each: watch it, change it, or handle it itself. The video’s example, Savvy Progress, uses all three: it watches subagent starts and model calls, changes the view above the prompt, and draws the whole agents panel itself.
Availability (per the video).
- Shipped on 1 October 2026 and switched on in the latest Claude Code. Needs 2.1.287 or newer.
- Works in the terminal and in the Code tab of the desktop app. At recording time it was not yet in the VS Code extension.
- No extra charge, but you need a paid plan (that is what includes Claude Code). A mod that calls a model in the background uses your normal usage.
- Most mods run on deterministic code, so they cost no tokens. The exception named is the watcher in #3, which runs a second agent.
🧩 The nine mods
| # | Mod | What it does | Cost / note |
|---|---|---|---|
| 1 | Savvy Progress | Progress bar above the prompt. For long runs with subagents, a live panel per helper showing its model, progress and rough API cost, with customisable pixel-claw mascots (designer, developer…) | Not stated (it reads events, so likely deterministic). The video’s showcase for watch / change / handle |
| 2 | Claude Skins | Reskins how tool calls render: icons per tool, coloured edits. The open-source repo ships seven themes (Tokyo is one), and you can ask for any palette or design system | Display only, so the safest first mod |
| 3 | ”You should know” | Built by Anthropic. A second agent reads Claude’s output and pops up the one thing you might miss in a wall of text. In the demo it flagged the cost of a video-generation call. You can steer what it watches (costs, customer issues…) | Extra tokens: it runs a second agent |
| 4 | File Tree | Right-hand file panel. A file shimmers while Claude touches it and turns green once the change is committed. Useful outside code too, e.g. which of many documents were opened or changed | No token cost |
| 5 | Cache Tax | The prompt cache goes cold after about an hour on a normal subscription, and the next message then re-reads the whole conversation. This mod stops you before such a message and shows the token / API cost. /keepwarm pings the session just before the cache expires | Can save usage in long sessions |
| 6 | Blast Radius | Built by Anthropic. Intercepts risky commands such as deleting a folder, lists exactly which files would go, and needs a human to press Proceed or Cancel | Guardrail, not a cost |
| 7 | Reflect | When your feedback contains trigger phrases (“no”, “this doesn’t look good”, “I wish we’d done it this way”), a band above the prompt offers to save it as a permanent rule in CLAUDE.md or a skill file. One click and Claude writes the rule. Jay likens it to RLHF on a personal, micro scale | Writes to your rule files |
| 8 | Terminal browser | Opens a real web browser beside the chat for terminal users, so you can review a page Claude built or a GitHub pull request without leaving the terminal | Terminal users only |
| 9 | Replay Theater | Anthropic sample mod. After a turn a replay button steps through every file edit Claude just made. Not limited to code: with an article it shows where Claude changed your draft | Not stated |
🧩 Build your own
- Update Claude Code to 2.1.287+ (
claude updatein the terminal; the desktop app shows an update popup). - Open Claude Code and describe the modification you want. A built-in skill,
plugin-authoring, tells Claude how mods are built. - It offers to turn on hot reloading for the session, so the mod appears as soon as Claude finishes building it.
- Iterate by talking to it normally (“make the bar thinner”).
- A mod Claude builds is temporary by default. When you like one, ask Claude to save it as a plugin and install it, and it loads every time.
đź“’ What I take from it
- Corroborated here: my Claude Code install lists a built-in
plugin-authoringskill described as “Make a mod”, which matches the name and purpose the video gives. Everything else about the nine mods is from the video alone; none were run while writing this note. - Ideas worth testing for this vault, not conclusions:
- Blast Radius is the UI version of my existing safety rule (“never delete user-authored content without confirmation”) and of the AI-Native SDLC Playbook line skills advise, hooks enforce, humans approve at gates. Mods are a third enforcement surface next to hooks.
- Reflect overlaps with Autoharness: both turn session corrections into durable rules. Reflect asks a human to click save each time; Autoharness distills and prunes automatically. The click-to-approve model is the safer end of that range, given DELEGATE-52’s warning that unattended edits compound.
- Cache Tax and
/keepwarmfit the cost levers in Token Optimization for Claude Code. - Terminal browser is redundant in the desktop app, which already has a browser pane.
- Malleable software: Jay hopes OpenAI follows and calls this the start of “more malleable software”. oh-my-pi takes a similar line from the other side: its extensions are plain TypeScript modules (“ask omp to write the piece you’re missing, then
/reload-plugins”). - Open questions: how mods are sandboxed, what they can read (a mod sees every tool event), and whether a third-party mod can be trusted. Treat installing one like installing a hook: read the code first. Relevant to AI Agent Security.
đź“– Takeaways
- A mod changes the harness itself, where a skill changes the instructions and MCP adds reach. That makes it the most invasive extension point, and the most powerful one.
- Start with a display-only mod (Claude Skins) before anything that sits on the event stream.
- The guardrail mods (Blast Radius, Cache Tax) are the ones with measurable value: they stop a costly or irreversible step before it happens.
- Ask Claude for the mod, keep it local while testing, and save it as a plugin only once it earns its place.
đź“– Further reading/watching
- Source video: 9 NEW Claude Mods that can truly change how you work (Jay E | RoboNuggets, 2026-10-08). The PDF with all nine prompts is linked in the video description (not fetched).
- Claude Code · Claude Code Best Practice · Harness Engineering · Agent Skills
- Extending Claude Code — Tools for Its Blind Spots — the skills and MCP tools I use to extend Claude Code
- Autoharness · Token Optimization for Claude Code · AI-Native SDLC Playbook · AI Agent Security
- oh-my-pi — another harness built around user-written extensions
- Meme Generation for Coding Agents — mods as the in-process host for a meme skill (
turn.complete, the Desktop-onlySvgelement); the mod API churns and mods are unsandboxed
Template: knowledge_note_info