Claude Code Mods: Which Are Worth Installing, and What They Cost

Every tool I have ever loved got worse the week its plugin marketplace opened. Not because plugins are bad, but because each one is a small, permanent decision about what may run inside the thing you depend on, and nobody makes those decisions carefully at the rate they install.

Claude Code now has mods. In a video published October 21, 2026, Ben of Ben AI walks through nine he says he kept after trying more than a hundred. It is a good tour. My question is narrower: which of them earn a place in a one-person build loop, what each one costs in access, usage and attention, and how to check a mod before it runs as you.

What a Claude Code mod actually is

A Claude Code mod is a plugin of JavaScript or TypeScript event handlers that runs inside Claude Code and can watch, change or take over what happens in a session.

Ben’s explanation is the right mental model: Claude Code emits a signal when something happens, and a mod listens for one. Anthropic’s mods overview names those signals precisely: prompt.submit when you send a prompt, tool.call before a tool runs, agent.spawn when a subagent starts, turn.complete when a reply ends, ui.render when part of the screen is drawn. The handler then does one of three things. It observes and passes the event on. It rewrites the event and passes on the changed copy. Or it answers the event itself, and nothing downstream runs.

That third move is what separates a mod from a Claude Code skill or a status line. A skill is instructions Claude reads. A mod is code standing in the doorway. It can hold your prompt, refuse a command, draw a pane, or send a request to a different model. The documentation lists the requirements plainly: Claude Code v2.1.287 or later in the terminal, and it works in the CLI, the Code tab of the desktop app and the VS Code extension.

Making one is now a sentence. Ask Claude for a mod and it uses a built-in plugin-authoring skill to write it. Per Anthropic’s create guide, a mod Claude writes lives in a per-session folder, asks your approval before it loads, and is deleted when that session ages out unless you copy it somewhere permanent. That is a sensible default, and the reason I would rather write a narrow mod than install a broad one.

The nine mods on one screen

The nine mods in Ben’s video are UI and workflow additions that either watch Claude Code or step into it, and the difference between those two groups is the whole buying decision.

Briefly, as Ben presents them. Brain dump catches a long voice prompt, asks clarifying questions and submits a structured version. Helper desk adds a progress band and a pane with one row per subagent and its cost. Cache clock counts down until the chat’s cache goes cold and can keep it warm. Handoff writes a detailed prompt for a fresh chat. Reflect spots a rule you gave Claude and offers to save it. Chat inventory lists the tools, connectors, skills and files a chat used. Screen share mode grays out names, emails, phone numbers and amounts. Replay theater plays back each file edit. Switchboard lists running mods with on/off switches.

Eight are Ben’s own builds or rebuilds, and he credits other community authors for the ideas behind a few. Replay theater is one of Anthropic’s published sample mods, which matters later: its source is public. Switch them on below and watch what accumulates.

The meter under the terminal is the point of the exhibit. With all nine on, the screen carries a band, a side pane with two sections, several chips and a toast, four mods can spend usage, and two of them sit in front of every prompt you type. Each was reasonable on its own. Together they are a different product from the one you chose.

The real price of a mod is reach, not dollars

A mod’s cost is everything it can see and change, plus the usage and attention it consumes, and the first of those is the one people skip.

Anthropic is unusually direct about this. The overview’s trust section says a mod runs with your permissions and is not sandboxed. Once loaded, it can read and write files anywhere your account can, read environment variables and settings files where API keys often live, see every prompt and tool call, rewrite either, submit prompts as if you typed them, call a model on your plan, and, unless your organization manages the setup, approve a tool call before you are asked. Turning on Claude Code’s sandbox does not contain it: the sandbox isolates the Bash commands Claude runs, and a process a mod starts runs outside it. The one surface a mod cannot restyle is the permission prompt.

Usage is the second cost. Ben splits mods into two types: those that use AI, and so spend your plan like a normal message, and those that only track or draw, which cost nothing to run. That is the right split. Brain dump and Handoff are in the first group by design. Reflect is unclear: it checks every prompt for a rule, and the video does not say whether that check is a model call. If it is, every message carries a small surcharge.

Attention is the third, and the one I care most about as a product person. In enterprise product work I watched dashboards multiply until nobody could say which number mattered. A progress bar you glance at during a twenty-minute agent run is useful. A progress bar, a countdown, a token counter and two panes during a two-minute edit are a cockpit you fly instead of the work.

The scores are mine and the scale is crude, but the shape is honest. The mods I would install cluster where reach is low: they watch events and draw. The ones I would skip cluster where they sit in the prompt path or write files. Handoff is the interesting exception: high reach, high value, worth it if you read what it writes.

Cache clock is the one with money in it

Prompt caching stores a copy of the conversation so the next request can reuse it cheaply, and when the cache expires, the next message pays to write the whole thing again.

Ben’s numbers check out against Anthropic’s prompt caching page. On Claude Opus 5.5, input is $4 per million tokens, a cache hit costs 0.05× that, and a cache write costs 1.25× with the default 5-minute lifetime or 2× with the 1-hour option. On a 200,000-token chat that is about 4 cents to come back warm and $1.60 to come back cold on the 1-hour cache, exactly the pair he quotes, and a 40× gap. The 5-minute default applies to API calls; that a subscription session runs on a one-hour timer is Ben’s claim, and I have not found it stated in Anthropic’s documentation.

Play with the time-away slider and the keep-warm line tells you something the video does not dwell on. On a one-hour cache, keeping a chat warm over lunch is a ping or two and a bargain. On a 5-minute cache it is two dozen pings in two hours, and the saving over one cold restart nearly disappears. The clock is worth installing for the information alone. Keep-warm is a decision per chat, not a default.

The bigger lever is not keeping long chats warm. It is ending them. Ben pairs Cache clock with Handoff for that reason: past roughly 250,000 tokens, by his rule of thumb, a fresh chat with a good handoff prompt beats a compacted old one on both quality and cost. I already work this way on Ballistica, my game project, where each session ends with a written report the next session starts from. A mod that drafts that report is useful. A mod that writes it without me reading it is how a wrong assumption becomes the next session’s ground truth. See the token expenditure index for why spend per accepted result, not per message, is the number to watch.

Which I would install, and which I would skip

Install the mods that help you review what Claude did; be slow with the ones that change what Claude is told.

Replay theater, first. Reviewing edits is the bottleneck in every agent workflow I run, and stepping through them one at a time is better than reading a forty-file diff. It is Anthropic’s sample, so the source is published and reviewable. It only replays file edits, not text in the chat.

Chat inventory and Helper desk. Both are observe-only by description and cost nothing to run. On multi-agent runs, where I have had several lanes working in parallel, “which subagent is stuck and what has it read” is the question I ask most. These answer it at a glance. Close the panes on short tasks.

Screen share mode, with care. If you record or demo, it is a real improvement on inventing fake client names. But it masks what Claude Code draws, not what the model already read, not your notifications, not your other windows. Treat it as a mask. The data you never collect cannot leak: a demo project with demo data is still the safer recording.

Switchboard, later. Ben’s point that mods will sprawl like skills did is correct. Until you run more than three, the Installed tab in /plugin already lists and disables them, and claude --safe-mode turns every installed mod off for a session. A mod with power over your other mods is also the one to vet hardest.

Brain dump and Reflect, skip as packaged. Both sit in front of every prompt. Brain dump can hold one back entirely; if it misfires you are debugging silence. Reflect writes to memory. My project’s CLAUDE.md is curated on purpose: its rules each came from a specific incident, and the file loads on every turn. A rule promoted because I was irritated once in a chat is context I pay for forever. I wrote about that discipline in Harness²: capture corrections automatically, promote them deliberately. And four rules in one file shows how much a short, curated instruction file can carry.

How to vet a mod before it runs as you

Vetting a mod means reading what it asks Claude Code to do before you load it, then trying it for one session with an exit ready.

Anthropic built the instrument for this. Because a mod can only reach outside itself through the mods API, claude plugin validate ./the-mod can list, without running anything, every event the mod hooks and every API call it makes. The hooks: line tells you what it sees. The calls: line tells you what it does. Paste one below.

The flag I care most about is the combination: a mod that can read files, environment variables or settings and can also make network requests can send your keys somewhere. That is not proof of bad intent; a weather widget needs the network. It is a question the author should be able to answer in one sentence. This is the same exposure as prompt injection from the other side: content that should be data trying to become authority.

1. Get the source, not a binary. Ben distributes his mods as a zip through his resource page. Unzip it into its own folder and open it before anything loads it.
2. Run validate and read both lines. Ask why for every $.http, $.process, $.env and $.fs.write call, and for any tool.check hook.
3. Try it for one session. claude --plugin-dir ./the-mod loads it without installing.
4. Know the exits. Disable it in /plugin, start with --safe-mode, or set disableAllHooks in user settings.
5. Prune monthly. Delete any mod you have not consciously used in thirty days.

In the enterprise work I did for years, nothing with this much reach went into a production tool without a security review and an owner. A solo builder does not need the committee. You still need the review, and you are the owner.

Build the narrow one yourself

The best mod for a solo builder is often a ten-line one you asked Claude to write, because you know exactly what it touches.

The product-leader instinct here is build versus buy, and mods tilt it hard toward build. Ben’s Reflect does three things: detect a rule, ask where to save it, write the file. I would want only the first: a line under the reply saying “that sounded like a rule,” with no file write and no model call. That is a narrower, lower-reach mod, and describing it takes one sentence. The same goes for the cache clock: the countdown is the valuable part, and it needs no prompt-sending power at all.

The agent harness argument applies directly. Every rule, skill and mod you add is part of the system that shapes the agent’s output. Add them one at a time, know why each is there, and remove the ones that stop earning their place. Pick yours below and take the checklist with you.

# Claude Code mod kit

## Before anything
[ ] claude --version  (installed mods need v2.1.287 or later)
[ ] Keep each mod in its own folder. Never install from a zip you have not opened.
[ ] claude plugin validate ./<mod>  then read the hooks: and calls: lines
[ ] Reject any mod that reads files, env or settings AND makes network requests, unless the author explains why
[ ] Try each one for a session first: claude --plugin-dir ./<mod>

## Helper desk  (Ben’s rebuild of a community mod)
[ ] Source: Ben AI’s free resource pack (linked from the video). Validate before loading.
[ ] Expect: A progress band above the prompt, and a pane with one row per subagent: colour, progress, cost.
[ ] Cost: No model calls. One more pane competing for your attention
[ ] Watch: Close the pane on short tasks. A progress bar you stare at is still a distraction.

## Cache clock  (Ben’s build, inspired by a community mod)
[ ] Source: Ben AI’s free resource pack (linked from the video). Validate before loading.
[ ] Expect: Counts down until the chat’s prompt cache goes cold. Optional keep-warm pings before it expires.
[ ] Cost: Display is free. Each keep-warm ping is a real, small request
[ ] Watch: Keep-warm sends messages as you. Know when it stops.
[ ] Your cache timer: Ben says 1 hour on a subscription. Turn keep-warm on per chat.

## Chat inventory  (Ben’s build)
[ ] Source: Ben AI’s free resource pack (linked from the video). Validate before loading.
[ ] Expect: A pane that lists the to-dos, connectors, skills and files the chat has used, with counts.
[ ] Cost: No model calls
[ ] Watch: Observe-only should stay observe-only. Check that it returns next(e) untouched.

## Replay theater  (Anthropic sample mod)
[ ] Source: github.com/anthropics/claude-code-playground, claude-code/mods/replay-theater (shared as-is, no support)
[ ] Expect: /replay steps through the file edits Claude made in the last turn, one at a time.
[ ] Cost: No model calls
[ ] Watch: Files only. Text Claude wrote in the chat is not replayed.

## Kill switches
[ ] One mod: /plugin, Installed tab, disable it
[ ] Everything I installed, this session: claude --safe-mode
[ ] Everything I installed, every session: "disableAllHooks": true in ~/.claude/settings.json (also stops settings hooks and a custom status line)
[ ] Review monthly: delete any mod I have not consciously used in 30 days

A mod is a hire, not an app. Check its references, give it the smallest desk that does the job, and know how to let it go.

Frequently asked questions

Sources and limits

Researched October 22, 2026, the day after the video. The descriptions of Ben’s nine mods come from his video; I have not installed his resource pack or read its source, so which events and calls each mod uses are inferred from Anthropic’s documentation of what such a feature requires. Mod capabilities, version requirements, commands and cache pricing are from Anthropic’s documentation as of that date and may change between releases. Reach and value scores are editorial.

1. Ben AI — 9 Insane Claude Mods I Already Can’t Live Without (YouTube, October 21, 2026)
2. Claude Code Docs — Mods overview
3. Claude Code Docs — Create a mod
4. Claude Code Docs — Mods reference
5. Claude Docs — Prompt caching (TTL and pricing)
6. anthropics/claude-code-playground — sample mods