The best Claude Code mods: 18 worth installing, checked on 2.1.287

How this list was made

If you are new to the idea, the explainer covers what a Claude Code mod is, how it differs from a hook, and the trust model. The short form: a mod is a plugin whose hooks/hooks.json names a JavaScript or TypeScript module that Claude Code runs inside its own process, on its events, with your permissions. It can draw panes and bands, rewrite tool calls and prompts, and add commands. It is not sandboxed.

Four rules decided what is here:

• Public code. A demo video with no repository is not a mod you can install, so it is not on this page, however good it looked.
• Loads on the launch build. Each repository was cloned and its plugin directory run through claude plugin validate on 2.1.287. Mods that fail are named at the end rather than listed.
• Does one thing you would miss. Reach is a cost. A mod that lists $.process.run or $.http.fetch in its calls: line has to earn it.
• Someone is behind it. The author is named on every entry, and the install line is theirs, taken from their README or marketplace file.

Two conventions, so the entries stay short. Install lines are what you type at the Claude Code prompt; the same commands work in your shell with claude plugin in front. The part after the @ is the marketplace's name, not the repository's, which is why it sometimes looks odd. And when a README tells you to set CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1, skip it: that was the early-access flag, and 2.1.287 ignores it.

Agent visibility and progress

The strongest category, and the one that makes the biggest difference to how a long task feels. Each of these turns a spinner and a scroll of commands into something you can read at a glance.

What's Agent Doing, by Tzafrir Rehan

What it does: a small box above the prompt that always says what Claude is doing right now, in plain English, with how long it has been at it: "Find the commit that broke fixed-amount coupons · 1m 48s". Click the arrow and it expands into every step of the turn with its duration and outcome.

Why it stands out: it answers the question you actually have during a bisect or a long test run, which is "is it stuck or just slow", and it does it by describing the purpose of each command rather than echoing the command. It also knows the states that matter: waiting for your approval, a subagent at work, compacting, and idle with a summary of the last turn. Of the dozen "live status" mods published this week, this is the one with the most thought in its state list.

github.com/tzafrir/whats-agent-doing · MIT

plan-progress, by Kirill Serditov

What it does: live progress bars above the prompt. Claude breaks a medium or large task into stages and steps, and you watch each bar fill in real time, with the subagents working on each task shown as strips under its bar, their model and effort and the tool in use beside them. Soft sounds when Claude needs a decision, hits an error or finishes.

Why it stands out: several tasks side by side, each with its own bar, is the "pizza tracker" view people have been asking for, and the details are right: four states in one brightness, a plan that can be rewritten mid-run without losing finished steps, bars that are saved per session and come back on resume, and a /progress-demo command so you can see it before you trust it. Its README has a 14-second video with sound.

github.com/zycck/claude-mods · MIT

agent-flow, by Charles Tao

What it does: /flow opens a live tree of the session's subagents and teammates beside the transcript: status, type, elapsed time, the tool call each one is running and how long it has run, who is waiting for your approval, call counts, and tokens once an agent finishes. Where no pane can be drawn, it prints the same tree as text.

Why it stands out: "subagent sprawl" is a real complaint, and this is the read-only answer to it. Agents waiting for approval are highlighted, a tool call over 30 seconds and an agent silent for two minutes are marked, and loops the engine never listed appear under a collapsed group instead of vanishing. The tree is built from engine events and reconciled with $.agent.list() every two seconds while anything runs, which is the honest way to do it.

github.com/Charlie0113-T/claude-agent-flow · Apache-2.0

prompt-rail, by oikon48

What it does: a rail of every prompt in the session, drawn above the prompt or beside the transcript. Hover a mark to read that prompt and a summary of its turn, click to jump the transcript back to it.

Why it stands out: long sessions turn into a wall, and "where did I ask for that" is otherwise a scroll. The rail is navigation for the transcript, which nothing else provides, in two layouts. Mouse support works best with "tui": "fullscreen" in your settings. The README still describes the early-access flag; ignore that step.

github.com/oikon48/prompt-rail · MIT

Replay Theater, by Claude Code DevRel (Anthropic)

What it does: records every Edit and Write call during a turn, then /replay (or pressing r on the hint above the prompt) opens a pane that walks through the edits one diff at a time, with a strip of numbered steps and Prev, Next and Close buttons.

Why it stands out: a review view for how you actually read a change, file by file in the order it happened, and it never blocks or changes anything. It is one of Anthropic's three sample mods and the README says how it was built, so it doubles as the best worked example of a pane with buttons. Shared as-is, without support.

github.com/anthropics/claude-code-playground · Apache-2.0

Usage, limits and context

The most-built useful category, because a mod reads $.session.usage() directly and gets the context window, the 5-hour and weekly windows and cost from Claude Code itself, instead of estimating from a transcript the way the status-line tools do. These are the ones that go beyond a meter.

Token Weather, by Claude Code DevRel (Anthropic)

What it does: one line above the prompt with a weather icon for how full the context window is, the percentage, tokens used out of the window, a small chart of recent turns and how much the last turn added. Clear under 25%, Cloudy to 49%, Showers to 74%, Storm to 89%, and "Compact soon" from 90%.

Why it stands out: it is about 80 lines, it only watches and draws, and it is the mod the official getting-started guide builds step by step, so you can read every line of it and then change it. If you install one usage mod to learn from, this is it.

github.com/anthropics/claude-code-playground · Apache-2.0

usage-meter, by HolyGrail

What it does: context fill, 5-hour limit and weekly limit as a permanent readout above the prompt.

Why it stands out: it is built on session.measure, the event that fires after each turn and whenever a plan limit's percentage changes, so it updates when the number moves rather than on a timer, and it survives /clear, resume and compaction, which its hook list shows explicitly. The same marketplace carries three siblings worth a look: a notice board shared by every session in a repository, a PR relay that wakes the session when its pull request is merged or reviewed, and a zsh compatibility shim. The README is in Japanese; the install lines are not.

github.com/HolyGrail/claude-mods

idle-compact, by takahirom

What it does: compacts the conversation once after 50 idle minutes, while the one-hour prompt cache is still warm, so that coming back to a long session does not re-cache the whole context at full price.

Why it stands out: it is the most carefully reasoned small mod on the list. Only the main conversation arms the timer; subagent completions, aborted turns and failed turns are ignored. When the timer fires it re-checks the wall clock, and does nothing if more than 58 minutes have passed, because a Mac that slept has a cold cache anyway. It tells you in the transcript when it will compact, never re-arms, and never retries a refused compaction. Written against 2.1.280 and the README still mentions the flag; it validates clean on 2.1.287.

github.com/takahirom/takahirom-claude-code-marketplace · Apache-2.0

usage-wrapup, by Kevin Koch

What it does: watches your rate-limit windows, and when any of them has 5% or less left (configurable), every tool result in the main agent and in subagents carries a note telling the model to start no new work, finish or safely pause the current step, write a short handoff, and stop. The status line shows the percentage left and a toast fires once.

Why it stands out: it solves the actual failure mode of hitting a limit, which is not the interruption but the half-edited files and no handoff. Delivering the instruction through tool results is clever: it reaches subagents too, without touching the system prompt. Works only on a Claude subscription, where Claude Code has rate-limit data, and installs from a clone.

github.com/DaKev/usage-wrapup · MIT

cache-tax, by Karan Bansal

What it does: stops a cold send once with the estimated price of rewriting the prompt cache, and /keepwarm keeps the one-hour cache warm during a break with a small cache-shared fork per idle stretch.

Why it stands out: where idle-compact makes the context small before the cache goes cold, cache-tax keeps the cache warm or at least tells you the bill. Its README is the clearest explanation of the cache-write economics you will find, with a real recording of the refusal. These are API-equivalent costs; on a subscription they are not extra charges, which the README says too. Published during early access, so skip the flag step.

github.com/karanb192/cache-tax · MIT

Workflow and safety

Blast Radius, by Claude Code DevRel (Anthropic)

What it does: when Claude calls Bash with rm -rf, git reset --hard, git clean, a force push or a database migration, it holds the call, works out what the command would touch using the tools' own dry runs, and opens a pane listing it, with Proceed and Cancel buttons. Cancel sends Claude a refusal with the reason; Proceed runs the command as written. At narrow widths the same report draws above the prompt.

Why it stands out: it is the clearest demonstration of what a mod can do that a hook cannot: hold a tool call for as long as it takes you to decide, and show you the consequences while you do. The guide is honest about its limits, "it's a safety net, not a permission system", because it reads the command text, so a command hidden in $(...) or a script gets past it. Use permission rules for a hard block and this for the near misses.

github.com/anthropics/claude-code-playground · Apache-2.0

next-steps, by Thariq Shihipar

What it does: after each turn, suggests up to three next prompts above the input box. Press 1, 2 or 3 and that prompt is written into the box as a draft you edit and send yourself; 0 dismisses. The top suggestion also appears as the box's ghost text, so Tab takes it.

Why it stands out: it never submits a prompt on its own, which is the right call, and the suggestions can be slash commands and skills your session actually has, because the mod feeds $.command.list into the question. It uses $.model.fork, which shares the session's prompt cache, so each suggestion costs about one short reply. It is in Anthropic's community marketplace and is the one community mod on Anthropic's own showcase page.

github.com/anthropics/claude-plugins-community · MIT

cc-pr-tracker, by Seza Akgün

What it does: watches GitHub pull requests from inside the session: merge state, review state and required checks above the prompt, with alerts when any of them change. Paste a PR URL and it starts tracking it.

Why it stands out: the "is CI green yet" tab switch, gone. It runs gh for the data, so it lists a process call in its validation output and needs the GitHub CLI signed in; that is a reach it earns. Written for an early-access build and validates clean on 2.1.287; skip the flag step in its quick start.

github.com/sezaakgun/cc-pr-tracker · MIT

RemCTL, by Federico Viticci

What it does: Apple Reminders inside Claude Code. The plugin gives Claude RemCTL's tools and draws today's tasks above the prompt in your lists' colours, with a /reminders pane.

Why it stands out: it is the model for a mod that ships with a real product rather than standing alone. RemCTL is a Reminders CLI for the terminal and for agents, backed by a small signed macOS app that holds the Reminders permission; the mod is one plugin in that repository, which stands at 573 stars on GitHub as of 2 October 2026. The README states plainly that the band needs 2.1.287 and that older builds still get the tools. macOS only, by nature.

github.com/viticci/remctl · MIT

Fun, and things to do while you wait

The most numerous category, and the one that spread fastest during early access. Toys, but toys that cost zero tokens and show off what the band can draw.

cc-arcade, by Seza Akgün

What it does: games above the prompt: Snake, Tetris, 2048, Minesweeper, Flappy, Pong, a typing test, Space Invaders and Doom, plus a pet that grows as Claude tests, commits and edits. Games pause when Claude finishes a turn.

Why it stands out: the Tetris-in-Claude mod, and the pause-on-turn-end detail is what makes it usable rather than a distraction. It is also a good read on how a mod handles key presses through ui.press without ever reading the keyboard itself. Needs an interactive terminal.

github.com/sezaakgun/cc-arcade · MIT

Mindful Claude, by Anthony (halluton)

What it does: a guided breathing exercise above the prompt while Claude works: coherent breathing, box breathing, 4-7-8 and the physiological sigh, in four animation styles, and the spinner's word reads the breath. It appears when Claude starts and disappears when Claude answers.

Why it stands out: it reframes the wait instead of filling it, and it is tiny. Note that claude plugin validate on 2.1.287 warns that the name mindful-claude "reads as one of Anthropic's own"; it is a warning, not an error, and the mod loads.

github.com/halluton/Mindful-Claude · MIT

Storytime, by Maddie Dreese

What it does: a 260,000-parameter TinyStories transformer that runs inside the mod and writes a short story above the prompt about whatever Claude is working on, locally, with zero API tokens.

Why it stands out: a language model running inside a Claude Code mod, in plain JavaScript with no Node, is the most unexpected thing anyone has done with the API so far, and it reaches nothing outside the session. The stories are exactly as good as a 260K-parameter model's stories, which is the joke.

github.com/maddiedreese/cc-storytime · MIT

ambient, by Barış Demirhan

What it does: a living band above the prompt that Claude's work feeds, with sound: an aquarium, a bonsai, a skyline, constellations, weather that can follow the real sky, a train, Game of Life, a heartbeat, a fireplace, digital rain or a lofi desk. /ambient picks the scene.

Why it stands out: eleven scenes, each driven by tool calls and turns rather than a clock, and it is one of the few community mods whose README says it was built and tested on 2.1.287 itself. It ships a privacy policy, which a mod with network access for the weather scene should.

github.com/barisdemirhan/claude-ambient · MIT

Good mods that did not make the cut, and why

claude-queue and claude-mermaid, by Gal Elmalah. A stack above the prompt that holds prompts typed while Claude is working and sends them when the turn really ends, and Mermaid diagrams drawn inline in the transcript. Both are well made, and both fail claude plugin validate on 2.1.287 with the same error: a third party's plugin name may no longer start with claude-, because it "passes as one of Anthropic's own". They will be back the day they are renamed, and the repository is worth watching for that.

Everything with a video and no repository. A spinner that has a model draw cartoons of what the agent is doing, a multiplayer Doom server for people waiting on their Claude, an ASCII video player with sound: all posted this week, none installable. If they ship, they belong here.

The desktop pets and notch pets. A handful of mods launched on day one that pair a mod with a small native app to show every session's state outside the terminal. They are early, and showing many sessions outside the terminal is what CrewTower does for Claude Code, Codex and the other agents, so this page is not the place to rank them.

Finding more

• awesome-claude-code-mods, by Karan Bansal, scans GitHub for every mod and lists what each one can reach. As of 2 October 2026 it counts 359 mods in 373 candidate repositories, checked against 2.1.287, and says how many run processes, write files or reach the network. Start there when you want a mod this page does not have.
• claude.dev/mods is Anthropic's showcase, with the three samples, the built-in mods and the first community picks.
• Anthropic's directory accepts plugins that include mods; a listing there has had an automated scan and a human review before going live.
• The docs and the source of the built-in mods, which are complete plugins with tests.

Whatever you install, run claude plugin validate on a clone first and read the calls: line. A mod runs as you. The explainer has the full trust model, including the built-in guard on Team and Enterprise plans and the remote off switch.

A note on what mods do not cover

Every mod above draws inside the one session it is loaded in. When you run three Claude Code sessions and a Codex one, and the one that needs your approval is in a window behind the window you are reading, no band will tell you. That is the part CrewTower covers: a macOS app that shows every agent session in the MacBook notch and lets you approve, deny or answer from there. It uses settings hooks, not a mod, so it sits alongside anything on this list.

Questions

How do I install a Claude Code mod from this list?

Check that claude --version reports 2.1.287 or later. Then, in a Claude Code session, add the author's marketplace with /plugin marketplace add owner/repo, install the mod with /plugin install name@marketplace, and run /reload-plugins. The part after the @ is the marketplace's name field, which is not always the repository name; each entry on this page gives the exact line. To try a mod for one session without installing it, clone its repository and start Claude Code with --plugin-dir pointing at the plugin's directory.

Which Claude Code mod should I install first?

If you run long tasks, What's Agent Doing, which names the step Claude is on and how long it has been there. If you are on a subscription and watch your limits, Token Weather or usage-meter for the numbers and idle-compact to stop a cold cache from eating a window. If you let Claude run shell commands freely, Blast Radius, which holds rm -rf, git reset --hard and force pushes behind a Proceed button. All three are small, read-only in what they reach, and easy to remove.

Do these mods work in the Claude Desktop app?

The Code tab of the Claude Desktop app draws a mod's panes and bands once the app bundles Claude Code 2.1.287 or later, except for elements the docs mark terminal-only such as Raster and Image. Some of the mods here were built in the terminal only and say so in their README. None of them draw in the VS Code extension's chat panel, in claude -p, or in cloud sessions, where hooks run but nothing is drawn.

Why does a mod's README tell me to set CLAUDE_CODE_ENABLE_FUNCTION_HOOKS?

Because the mod was published during the public early access that ran for about four weeks before mods launched in Claude Code 2.1.287 on 1 October 2026, when the feature was called function hooks and needed that flag. Claude Code 2.1.287 and later ignores the variable; mods are on by default, and setting it to 0 does not turn them off. Skip that step and follow the rest of the README.

Are these Claude Code mods safe to install?

Every mod is code that runs with your permissions and is not sandboxed, so treat the list as a reading list, not an allowlist. All eighteen have public source and passed claude plugin validate on 2.1.287 on 2 October 2026, which lists the events each one hooks and the mods API calls it makes. Run it yourself on a clone before installing, and expect a usage meter to reach nothing beyond the session and the screen, while a PR tracker or a Reminders band will legitimately run processes or make network requests.

Can I run several Claude Code mods at the same time?

Yes. Mods form a chain: the first to load sees an event first and its result last, so mods from different authors stack. Two things to watch: the band above the prompt is shared by every mod, so several band mods stack vertically, and installed mods share one worker thread, so a mod that blocks the thread is unloaded and three untraceable crashes turn every non-built-in mod off for the session.