Capital & Compute

Best Claude Code Mods and Where to Find Them

The best Claude Code mods by job: usage-meter for plan limits, context-bar for context, whats-agent-doing for live activity, and how to vet the rest.

· Updated October 24, 2026· ai· coding-agents· tools· By Capital & Compute
The two Claude Code mod directories worth using; only the awesome list shows what each mod can reach.

The short answer. Five mods cover most needs: usage-meter for plan limits and session cost, context-bar for what fills the context window, whats-agent-doing for what Claude is doing right now, ci-watch for the CI run after a push, and transcript-redactor to keep API keys out of the transcript. The first three only watch and draw. For more, browse claude-mods.com and check any pick against the awesome-claude-code-mods catalogue, which scans every mod for what its code can reach. Run claude plugin validate on a mod before you install it. For jobs no settings hook can do, the standouts are debugger (pause and edit the agent loop), goalpost-guard (stops Claude weakening a test to pass it) and undo-agent (undoes a subagent’s edits, which /rewind does not).

Anthropic launched Claude Code mods on 1 October 2026 and described them as “small TypeScript functions that change how Claude Code works.” Mods ran in early access through September, so by launch day there were already more than 70 on GitHub. Eleven days later the awesome-claude-code-mods scanner counted 3,561 mods in 1,825 repositories; by 20 October it counted 5,667 in 2,811, and its 22 October scan counted 6,092 in 3,048. Most are experiments, duplicates or one-person tools. This page is the short list: 61 mods that each do one job, are maintained and licensed, and say what they touch.

The best Claude Code mods at a glance

Every mod here is free with its source on GitHub, and most install the same way: add its marketplace with /plugin marketplace add <owner>/<repo> (the linked repository), then run /plugin install <plugin ID>. “See repo” means the mod installs another way, explained in its section below. Reach is the widest thing a mod’s code can touch, from the awesome-claude-code-mods scan: L0 draws and remembers, L1 reads, L2 writes files, runs processes or drives Claude, L3 reaches the network. Stars and last push were read from the GitHub API on 24 October 2026, and reach from the catalogue’s 22 October scan.

Mod Job Reach Stars License Last push Plugin ID
token-ledger Cost and plan usage L0 10 None 16 Oct token-ledger@claude-mods
usage-meter Cost and plan usage L1 370 MIT 23 Oct usage-meter@claude-code-mods
usage-band Cost and plan usage L1 16 MIT 3 Oct usage-band@cc-usage-band
token-weather-usage Cost and plan usage L1 21 Apache-2.0 22 Oct token-weather-usage@augiefra-mods
cache-tax Cost and plan usage L2 249 MIT 22 Oct See repo
budget-guard Cost and plan usage L2 10 None 16 Oct budget-guard@claude-mods
cctop Cost and plan usage L2 16 MIT 1 Oct cctop
cache-keepalive Cost and plan usage Not scanned 0 MIT 15 Oct cache-keepalive@richardjsun-mods
context-bar Context window L1 370 MIT 23 Oct context-bar@claude-code-mods
context-report Context window L1 52 MIT 23 Oct See repo
context-xray Context window L1 11 MIT 9 Oct context-xray@claude-pro-kit
session-meter Context window L2 7 MIT 23 Oct session-meter@claude-code-mods
token-weather Context window Not scanned 284 Apache-2.0 14 Oct See repo
agent-flow What Claude is doing L0 25 Apache-2.0 9 Oct agent-flow@claude-agent-flow
flightdeck What Claude is doing L1 135 MIT 12 Oct flightdeck@claude-flightdeck
subagent-ledger What Claude is doing L1 12 MIT 4 Oct subagent-ledger@kilimcininkoroglu-mods
whats-agent-doing What Claude is doing L1 8 MIT 4 Oct whats-agent-doing@tzafrir
activity What Claude is doing L1 10 MIT 6 Oct activity@claude-mods
where-am-i What Claude is doing L2 370 MIT 23 Oct where-am-i@claude-code-mods
plan-progress What Claude is doing L3 58 MIT 20 Oct plan-progress@zycck-mods
downrange What Claude is doing L2 0 MIT 22 Oct downrange@downrange
review-watch Repo, PR and CI L2 370 MIT 23 Oct review-watch@claude-code-mods
worktree-map Repo, PR and CI L2 92 MIT 3 Oct worktree-map@awesome-claude-code-mods
ci-watch Repo, PR and CI L2 7 MIT 23 Oct ci-watch@claude-code-mods
cc-pr-tracker Repo, PR and CI L2 6 MIT 10 Oct cc-pr-tracker@cc-pr-tracker
pr-watch Repo, PR and CI L2 9 MIT 23 Oct pr-watch@claude-code-pr-watch
gh-status Repo, PR and CI Not scanned 6 MIT 13 Oct gh-status@oborchers
orphan-server Your machine L2 12 MIT 4 Oct orphan-server@kilimcininkoroglu-mods
collision-guard Safety and secrets L2 44 MIT 2 Oct collision-guard@nateherk-mods
git-gates Safety and secrets L2 0 MIT 23 Oct See repo
honmoon-redact Safety and secrets L3 6 Apache-2.0 19 Oct See repo
transcript-redactor Safety and secrets Not scanned 92 MIT 18 Oct transcript-redactor@claude-code-mods
secret-box Safety and secrets Not scanned 92 MIT 18 Oct secret-box@claude-code-mods
blast-radius Safety and secrets Not scanned 284 Apache-2.0 14 Oct See repo
image-view Reviewing and viewing L2 251 MIT 12 Oct image-view@claude-image-view
filetree Reviewing and viewing L2 277 MIT 23 Oct filetree@claude-code-filetree
prompt-rail Reviewing and viewing L2 45 MIT 23 Oct prompt-rail@oikon48
pinboard Reviewing and viewing L2 11 MIT 7 Oct pinboard@pinboard
replay-theater Reviewing and viewing Not scanned 284 Apache-2.0 14 Oct See repo
terminal-browser Reviewing and viewing L3 4,186 MIT 21 Oct See repo
cc-changelog Reviewing and viewing Not scanned 0 Not detected 23 Oct cc-changelog@cc-changelog
storybloq Memory and long sessions L1 781 Custom 19 Oct See repo
ctx-handoff Memory and long sessions L2 122 MIT 18 Oct ctx-handoff@ctx-handoff-mod
lcm Memory and long sessions L3 60 MIT 21 Oct lcm@lossless-claude
Mindful-Claude While you wait L0 108 MIT 14 Sep mindful-claude@mindful-claude
cc-arcade While you wait L0 44 MIT 15 Sep cc-arcade@cc-arcade
debugger Debug and audit L2 4 MIT 6 Oct debugger@supermods
drift-radar Debug and audit L2 0 MIT 6 Oct drift-radar@drift-radar
headspace Debug and audit L2 2 MIT 18 Oct headspace@emcc
absolutely-not Keeping Claude honest L1 0 MIT 6 Oct absolutely-not@absolutely-not
goalpost-guard Keeping Claude honest L2 2 MIT 18 Oct goalpost-guard@emcc
witness Keeping Claude honest Not scanned 0 MIT 19 Oct witness@iniphi-witness
undo-agent Undo L2 1 MIT 23 Oct undo-agent@modhub
better-rewind Undo L2 0 MIT 11 Oct better-rewind@better-rewind
quota-guard Unattended runs Not scanned 0 MIT 20 Oct See repo
taskcut Unattended runs L2 7 MIT 27 Sep taskcut@taskcut
mcp-diet Leaner requests L0 1 MIT 23 Oct mcp-diet@modhub
stay-put Leaner requests L1 0 MIT 4 Oct stay-put@ground-rules
codex Other agents and review L2 2 Apache-2.0 4 Oct codex@claude-mods
redpen Other agents and review L2 4 MIT 11 Oct redpen@redpen
dep-sentinel Other agents and review L3 12 MIT 4 Oct dep-sentinel@kilimcininkoroglu-mods

Four authors on this list named their marketplace claude-code-mods and three named theirs claude-mods, so the part after the @ does not identify the author on its own. Check which repository you added before you install.

What are Claude Code mods?

A mod is a plugin that changes how Claude Code looks and behaves, written as JavaScript or TypeScript event handlers. Claude Code fires an event each time it does something (calls a tool, takes a prompt, draws the spinner), and the mod’s handler can watch it, rewrite it, or answer it so the normal behavior never runs.

That last part is the new thing. In practice it means a mod can:

  • draw a pane beside the transcript or a band above the prompt, with buttons and text fields
  • redraw parts of Claude Code itself, such as a tool row or the spinner
  • hold a tool call, rewrite it, or refuse it
  • add a /command that runs your code instantly, with no Claude turn and no tokens

Mods need Claude Code v2.1.287 or later in the terminal (the Desktop app’s built-in copy supports them from v2.1.286) and are on by default, per Anthropic’s mods documentation. Hooks run everywhere Claude Code runs, including claude -p and cloud sessions. What a mod draws appears in the terminal, the Desktop app’s Code tab and, since extension v2.1.299, the VS Code extension’s chat panel; claude -p, the Agent SDK and cloud sessions run the hooks but draw nothing.

Some of Claude Code’s own features are now mods. /diff is one; so is AGENTS.md support, the plugin-authoring skill and the opt-in “You should know” side agent, which flags things you might have missed above the prompt. The source of four built-ins (diff, agents-md, sec-default, telemetry) is public in the mods folder of the Claude Code repository, and it is the best reference code there is if you plan to write your own.

Claude Code mods vs hooks: what changed?

Claude Code already had hooks: shell commands, HTTP requests or prompts you set in a settings file, run at fixed lifecycle points. Mods don’t replace them. The difference is where the code lives.

Mod Settings hook
Runs as Functions inside the Claude Code process A separate shell command, HTTP request or prompt
Keeps state between events Yes, in the module’s variables No, one process per call
Can draw in the interface Yes No
Can change Tool calls, prompts, commands, turns, the UI Whether a call goes ahead, its arguments and result, added context
Written in JavaScript or TypeScript Any script plus a settings.json entry

Order matters when both exist. A PreToolUse hook in managed settings runs first and its block is final; then mods run; then the PreToolUse hooks from other settings files and plugins, according to the admin documentation. So a mod that answers a tool call itself can stop your own project-level hooks from ever seeing it. If you rely on a hook as a guard, keep that in mind. (The Claude Code harness guide covers where settings hooks sit among skills, subagents and MCP.)

Where to find Claude Code mods

There is no official mods-only gallery yet. Anthropic points to the Claude directory and /plugin, where mods appear as plugins alongside everything else, and to its own official plugins repository, which does not separate mods out. Discovery has moved to community sites. These are the ones that existed on 2 October 2026, checked that day; the claude-mods.com and claudemods.ai counts were read again on 24 October 2026 and the awesome list count comes from its 22 October scan.

Directory Mods listed How entries get in Safety information Best for
claude-mods.com 293 (75 on 2 October) Daily GitHub scan, then each mod installed and checked by hand against its README; needs a license or a clear author Category only Browsing what exists
awesome-claude-code-mods 6,092 in 3,048 repos (72 on 2 October) Nightly scan of every repo with a hooks module, plus a short hand-picked Featured list Footprint per mod: events hooked, API calls, reach level Vetting a mod before install
claudemods.ai 45 in 9 categories Submissions, plus a daily vote for “mod of the day” None Light browsing
claude-code-mods (claudecodemods.com) Mixed with plugins, skills, agents and MCP Pull request; CI checks the source URL responds None, by its own statement Wider extension catalogue
claudemod.com 155 entries, mostly skills and MCP servers Community listings None Not mods specifically
Claude directory Not separated Anthropic’s plugin submission flow Plugin-level Official installs
Five mod directories against what you need from themComparison of five directories. Mods-only: claude-mods.com yes, awesome-claude-code-mods yes, claudemods.ai yes, claudecodemods.com mixed, claudemod.com mostly not. Hand-checked entries: claude-mods.com yes, awesome list scanned not tested, claudemods.ai no, claudecodemods.com URL check only, claudemod.com no. Shows what a mod can reach: only awesome-claude-code-mods. Size: 293 on 24 October, 6,092 on 22 October, then 45, mixed and 155 mixed.AdvantageTrade-offDrawbackDimensionclaude-mods.comawesome listclaudemods.aiclaudecodemodsclaudemod.comMods onlyYesYesYesMixedMostly notEntries checkedInstalled by handScannedNoURL onlyNoShows what it can reachNoPer modNoNoNoMods listed2936,09245Mixed155 mixed
Five mod directories against what you need from them
Dimensionclaude-mods.comawesome listclaudemods.aiclaudecodemodsclaudemod.com
Mods onlyYesYesYesMixedMostly not
Entries checkedInstalled by handScannedNoURL onlyNo
Shows what it can reachNoPer modNoNoNo
Mods listed2936,09245Mixed155 mixed
Only one directory tells you what a mod can touch before you install it. That is the one to check last, whichever you browse first.Source: Each directory's own page, read 2 October 2026; claude-mods.com count updated 24 October, awesome list from its 22 October 2026 scan

One trap: the GitHub repo that ranked first in search for “best Claude Code mods” on 2 October, 0xDarkMatter/claude-mods, is not about mods in this sense. It was created in November 2025 and bundles skills, agents, commands and settings hooks; the name collision is an accident of timing. It is a fine project. It is just not what you are looking for.

The best Claude Code mods, by job

Seeing what a session costs and how much plan is left

This is where mods earn their keep first, because Claude Code’s own /cost command only answers when you ask.

  1. token-ledger pins a status line with the session cost, the last turn’s tokens and the cache hit ratio, and adds /ledger for the full table. It is one of the few L0 mods on this list: it only draws and remembers. The repository has no license file, so read it before you depend on it.
  2. usage-meter shows your 5-hour and 7-day plan usage, the countdown to each reset and the session’s cost. It comes from a 200-star collection of small mods whose README shows each one in a screenshot. If you are on a Pro or Max plan, this is the first mod to install.
  3. usage-band fits the same job into one line: both plan windows with their resets, context tokens and the last turn’s cache hit. Its README says it reads no files, runs no processes and makes no network requests, and the scan agrees it stays at L1.
  4. token-weather-usage packs the most into one band: context, both plan windows against the clock, how long the prompt cache stays warm and what letting it lapse would cost, and what the session and last prompt cost.
  5. cache-tax warns before a send that would rewrite a cold cache, with the estimated cost, and /keepwarm 90m keeps the cache warm while you step away. The pings cost tokens, which the README says plainly. It installs from the author’s wider marketplace: /plugin marketplace add karanb192/claude-code-mods, then /plugin install cache-tax@claude-code-mods.
  6. budget-guard, from the token-ledger author, watches session dollars and both plan windows. At 80% of a limit it warns once; past the limit, by default, it refuses tool calls and asks before the next prompt. That is a hard stop a settings hook can’t draw or explain.
  7. cctop opens a btop-style pane with context fill, tokens, cost, cache hit ratio, rate limits, per-tool latency and subagents. More data, more reach: it watches every tool call and compaction.
  8. cache-keepalive pings the cache with a one-line fork every 50 minutes while a background task outlives it, and stops once the pings have cost what one rewrite would. The savings logic is the author’s own, so treat it as a claim.

None of these change the bill by themselves. They change when you notice it, which is the variable that matters (an earlier replay of 500 sessions found that token-saving tools cut a bill by far less than claimed, and the one prompt-level tool with a measured saving, Ponytail, cut cost per task by 10% to 26% in two tests).

Seeing what fills the context window

  1. context-bar draws the context window as one stacked bar, a color per category, with token counts and the point where it compacts. /context-bar opens it. Same collection as usage-meter.
  2. context-report prints which instruction files Claude Code loaded (tier, path, @ imports and size) once per context load, and warns when a project CLAUDE.md keeps its AGENTS.md from being read. That answers “is my memory file actually loading?” It ships inside Darkroom Engineering’s cc-settings bundle, so install it through that repository’s installer.
  3. context-xray opens /xray, an exact breakdown of what fills your context window: what is sent with every request, what loads on demand, and your memory files and listings. It is one of 23 mods in a kit built to make the Pro plan last longer.
  4. session-meter docks a pane with context by category, where auto-compact kicks in, context growth over turns, the three heaviest items, and both plan windows with whether your pace lasts until the reset. It writes files and can drive Claude, so it sits at L2.
  5. token-weather is Anthropic’s own one-line context forecast, covered with the other samples below.

Seeing what Claude is doing right now

  1. agent-flow draws a live tree of subagents and teammates with /flow. L0: it only draws.
  2. flightdeck is the fullest dashboard here: context and cost, an advisor timeline, every permission check, and subagents as cards or swimlanes. It is in the awesome list’s hand-picked Featured set, and the scan puts it at L1 despite how much it shows.
  3. subagent-ledger answers the question /cost can’t: which subagent spent what. One row per agent with its turns, wall-clock time and tokens, costliest first.
  4. whats-agent-doing names the step Claude is on in plain English, and how long it has been on it, in a box above the prompt. Click it to see every step so far. Display only: no network, no files.
  5. activity shows what Claude is doing this second: the command it is running and for how long, anything waiting for your approval, running subagents, and its todo list with progress. The scan lists a tool.check call, the hook that can approve tool calls, so read that part of the code.
  6. where-am-i keeps the session’s goal, the current step, what waits on you and the next step above the prompt, with /where. The scan shows it drives Claude to write that summary, so it can spend a little usage.
  7. plan-progress turns Claude’s plan into live progress bars. It is one of the editor’s daily picks, covered in detail below.
  8. downrange shows progress and an ETA for background shells and Monitor watches, read from their real output or past runs, with no model request. New, with no stars yet, but its README is unusually complete.

Watching the repo, pull requests and CI

  1. review-watch shows one live line per running code review, whether from Codex or a review subagent, with the model, target, elapsed time and latest output. A toast lists the findings when it ends.
  2. worktree-map lists every git worktree with its branch. An editor’s pick, covered below.
  3. ci-watch takes over after Claude pushes a branch, opens a pull request or pushes a tag. A band shows each GitHub Actions run with a progress bar timed against that workflow’s last successful run, then the result in green or red, also as a toast.
  4. cc-pr-tracker shows watched pull requests above the prompt, with merge state, review and required checks.
  5. pr-watch keeps a pane of the pull requests the session touched: CI, review, conflicts and merge. You see the moment the pull request Claude opened goes green, gets a review or hits a conflict.
  6. gh-status puts the current branch’s pull request in one line (state, verdict, pipeline, reviews, unresolved threads and the GitHub account in use), and /pr opens the detail. New and not yet scanned. It works through your gh login (a separate one per repository if you want), so expect network reach.

All of the pull-request mods need GitHub access through gh or the API. That reach fits the job.

Keeping your machine clean

orphan-server fixes a small, maddening problem. Claude starts a dev server in the background, the session ends, and days later you get address already in use. This mod lists servers that a Claude Bash call started in the current repository and left listening on a port, with their age and a stop button for each. It checks with lsof at session start, after each turn and every 60 seconds.

Keeping secrets out of the model and stopping bad commands

  1. collision-guard asks before Claude edits a file that another open chat changed in the last 30 minutes. If you run several sessions in one project, this is the guard to add first. Its README says the mods in that repository make no network calls of their own.
  2. git-gates refuses a commit, push or merge unless your latest message said so. It now sits in its author’s MIT-licensed skills repository and installs with that bundle.
  3. honmoon-redact redacts API keys and sensitive identifiers from Read, Bash and Grep output. It wraps the honmoon command-line tool, which you install first.
  4. transcript-redactor swaps every key that reaches the conversation for a stable placeholder, whether you pasted it, a tool printed it, or it sits in CLAUDE.md. The real value goes back in when a command runs, so commands still work. The vault lives in memory and is never written to disk. It replaces secret-redactor, which an earlier version of this article listed; the author’s repository now carries this mod instead.
  5. secret-box, from the same author, adds a masked input box: when Claude needs a key, you paste it there, the mod writes it to .env, and Claude only ever sees a placeholder. It refuses to commit .env. It includes transcript-redactor, so run one or the other.
  6. blast-radius holds risky shell commands until you approve them. One of Anthropic’s samples, below.

A redactor is the clearest case for a mod over a hook: it has to read every tool result and every prompt, which is a lot of reach, granted for a narrow job. Read the code.

Reviewing what changed and seeing more in the terminal

  1. image-view shows thumbnails of the images you paste above the prompt, instead of bare [Image #1] tags.
  2. filetree draws an IDE-style file tree with Git status per file and highlights the files Claude reads or edits as it works. It is in the awesome list’s hand-picked Featured set.
  3. prompt-rail keeps a rail of your past prompts: hover to reread one, click to jump back to it.
  4. pinboard pins open decisions, the task list and links Claude creates in a pane, so they don’t scroll out of view.
  5. replay-theater steps through the last turn’s file edits one diff at a time. Anthropic’s sample, below.
  6. terminal-browser and cc-changelog are editor’s picks with full write-ups below.

Keeping memory across long sessions

storybloq (781 stars) keeps tickets, issues, handovers and lessons in a project folder and works across Claude Code and Codex; install it with its own setup command, which adds the mod. ctx-handoff is a long-session toolkit: it keeps sessions going on their own, has Claude remember what you taught it, and catches its usual slips. lcm stores every message and compacts old turns into a tree of summaries, so nothing is lost outright. All three see every tool call, and ctx-handoff and lcm also see the system prompt. These are the heaviest mods on this list, so they come last.

While you wait

cc-arcade runs nine games above the prompt (Snake, Tetris, 2048, Minesweeper and more) and pauses them when Claude finishes. Mindful-Claude draws a guided breathing band while a turn runs, with four timings. Both are L0, harmless in reach terms, and the fastest way to see what a mod can draw.

Which Claude Code mods do things nothing else can?

Most mods above draw something a status line could almost draw. The 15 here use the parts of the mods API that settings hooks and skills never had: rewriting a section of the system prompt, changing what Claude is told about a tool, holding a model response before Claude acts on it, compacting on a mod’s own schedule, or standing in for a subagent. Each fills a gap in a built-in feature, named below. All are under a month old, so their star counts are low. Read the reach before you install.

Seeing and steering the agent loop

  1. debugger is a step debugger for Claude Code itself. It pauses at a prompt, a model request, a model response, a tool call or a tool result, and while an event is held you can edit it: the response text before Claude reads it, a tool argument, or the model and effort of one request. Then step, continue or stop the turn. Breakpoints can follow Claude into subagents. Its README lists no network and no file access.
  2. drift-radar fingerprints the system prompt sections, the built-in tool descriptions, the model and the effort each session, and toasts when any of them changed since your last session on the same model and mode. /drift shows the diff, section by section, including changes no changelog lists. When Claude Code feels worse after an update, this is your own evidence of what changed in what the model was given.
  3. headspace shows what fills the context window, not just how full it is: system, rules, conversation and large tool results. Then you can take one piece out, such as a single CLAUDE.md rule for this session or a 40,000-token cat from an hour ago, with no summary and no model call. The pane says what each removal costs in prompt cache. It installs a second plugin, headspace-cut, to do the removal.

Keeping Claude from gaming the task

  1. goalpost-guard catches Claude making a check pass by lowering it: a deleted or rewritten assertion, an added .skip or .only, @ts-ignore, strict turned off, continue-on-error on a CI step, a lowered coverage threshold, a snapshot rewrite with -u, or a commit with --no-verify. It holds the edit or command and asks you, and whatever you type goes back to Claude as the reason. Its README shows a recorded session where Claude, blocked, fixes the code instead of the test.
  2. witness keeps a list of what you asked for and counts an item done only when a later read-back proves it (a file read, a SELECT, a fetch), so Claude cannot certify its own write. It also holds an execute_sql UPDATE or DELETE that has no RETURNING clause. It is the heaviest mod in this section: Haiku side passes read your prompts, at a median of about 52,000 tokens a session by its own measurement, and the scan had not reached it yet.
  3. absolutely-not adds a short tone section to the system prompt (no “You’re absolutely right!”, say so when you are wrong), strips the praise opener from each stored reply so the model never reads its own tic back, and /tics counts the ones that still get through.

Undoing what /rewind can’t

/rewind restores your own turns’ file edits. Anthropic’s checkpointing documentation lists what it misses: changes made by Bash commands, and most subagent edits, including a background forked skill and /code-review --fix (“use git to revert them”).

  1. undo-agent snapshots each file before a subagent edits it, and /undo-agent last puts the files back after a confirmation (files the agent created are removed). Your own turns are left to /rewind.
  2. better-rewind replaces /rewind with a list that also holds every tool call and every question Claude asked you. Rewind to just before one tool call, or go back to a question and pick a different answer, and Claude carries on from there. It previews the reply that followed each point and keeps the branch you rewound away, readable with /better-rewind branches. Needs Node.js on your PATH.

Running while you are away

  1. quota-guard watches the 5-hour and weekly windows. Near the limit (94% by default) it finishes the turn, compacts, pauses until the reported reset, and then resumes the task by itself, surviving restarts and sleep. It also compacts an idle session just before the prompt cache expires, so the next request doesn’t resend everything at full price. It loads from a cloned folder with --plugin-dir or CLAUDE_CODE_PLUGIN_DIRS, not from a marketplace.
  2. taskcut changes when Claude Code compacts. Hand it “do A, B and C” and walk away: once the context is past a floor, it compacts between finished pieces of work instead of halfway through C when the window fills. The compaction is Claude Code’s own /compact. The judge is one Sonnet call per step past the floor, which its README puts at about half a cent.

Cutting what every request carries

  1. mcp-diet sets each MCP tool, whole MCP server or subagent type to on, defer (Claude sees only the name until it searches) or hide, per project, with each description’s size. Tool search already defers tools across the board; this lets you say “in this repo, never the Jira tools.” Changes apply from the next session. It is L0.
  2. stay-put refuses cd dir && … and hands Claude the one-command form (git -C api log, pnpm --dir web). A chained cd stops a command from matching the permission rules you wrote for the plain command, so this cuts prompts as well as noise. Add the marketplace with /plugin marketplace add ivanvyd/ground-rules.

Other agents, line-by-line review and packages

  1. codex makes OpenAI Codex a native Claude Code subagent. Claude starts it with the Agent tool as codex:read or codex:write, it shows in the agent list with its model and token count, and you can message it while it runs. No Claude model runs inside: the mod drives codex app-server. It needs the Codex CLI installed and logged in, and macOS or Linux.
  2. redpen opens Claude’s last reply, or the diff of any file it changed, line by line. Comment on a line or a range, then send every comment back as one prompt, the way a pull request review works. Some list labels are in Korean; the commands and keys are not.
  3. dep-sentinel checks each package Claude installs against its registry and OSV.dev, and stops one that doesn’t exist, is brand new, looks like a better-known name, is outdated or has a known vulnerability, naming the latest version. That is the check for a hallucinated or typosquatted name, made before the install runs rather than after. It comes from the orphan-server author’s marketplace.

What are Anthropic’s sample Claude Code mods?

Anthropic publishes three small, complete mods in its claude-code-playground repository, each with a README that explains how it was built. They are the best code to read before you write your own, and two of them are useful as they stand.

  • token-weather draws a forecast of your context window in the band above the prompt, such as “Showers, 67% of context”, with a sparkline of recent turns. It reads real figures from the session’s usage, the same numbers as the status line, not an estimate. It is the smallest of the three: one hooks module and one line of UI.
  • blast-radius holds a risky shell command (rm -rf, git reset --hard, a force push, a database migration), works out what it would change, and opens a pane with Proceed and Cancel. If you cancel, Claude sees the refusal and the reason. It runs commands to measure, so it reaches further than the other two.
  • replay-theater records each file edit Claude makes in a turn and, when the turn ends, lets you step through them one diff at a time. It only watches: it never blocks, changes or delays an edit.

To install one, clone the repository and follow the folder’s README. They load with claude --plugin-dir ./token-weather for a single session, or install from a local marketplace named claude-code-playground-mods.

Which Claude Code mods are worth using every day?

These four are the editor’s own picks for day-to-day development work. The list above orders mods by how little they reach; this one is chosen on usefulness alone, so check each against the safety steps below before you install it. Repository details were first read from the GitHub API on 4 October 2026 and star counts refreshed on 24 October 2026. For mods aimed at watching a session (live activity, context, CI, memory and plan usage), see Claude Code dashboard mods.

plan-progress: see a long task move

plan-progress turns Claude’s plan into a live progress bar above the prompt: one row per task with its stage, percent and time, and a strip under it for each subagent showing its model, effort and current tool. Approved plans become bars, bars survive a resumed session, and short sounds mark a question, an error and completion. /progress shows or hides it. Its claude-mods.com listing says it makes no network calls, but the 20 October release added one: pressing the version number in its settings asks GitHub for a newer release and can install it, which moves it to L3 in the catalogue’s 22 October scan. Nothing is fetched unless you press it. MIT, 58 stars.

plan-progress shows two tasks above the Claude Code prompt at 56% and 75%, each with subagent rows naming model, tool and time.
Pinch or scroll to explore
Two tasks in flight, each with its subagents listed under the bar, and a live percent on the right. Image from the plan-progress repository (MIT).Source: zycck/claude-mods

cc-changelog: know what changed in Claude Code

cc-changelog opens the unofficial Claude Code changelog in a pane with /whatsnew: the 25 latest releases, search, and single entries. It also gives the model fifteen tools, so Claude can look up a flag, an environment variable or the mods API itself, and it checks for new releases every 30 minutes with a toast. Its claude-mods.com listing notes the reach: it calls a third-party changelog site (changelogs.core-directive.com) and runs version checks. Listed as MIT, though GitHub does not detect a license file.

Illustration of a terminal with a highlighted cc-changelog pane drawn above the Claude Code prompt.
cc-changelog adds a pane to the session; Claude Code moves fast enough that a changelog one command away earns its place. Illustration from claude-mods.com.Source: claude-mods.com

terminal-browser: open a page without leaving the session

terminal-browser puts a Chromium browser in a pane beside the transcript. Point it at a pull request, a docs page or your local dev server with /browser <url>, and the agent can browse it too. It is the most-starred mod in this article at 4,186 stars. It installs as its own program (brew install terminal-browser, or the script on its README), not through /plugin. MIT. It runs a full browser, so treat it as one of the widest-reaching mods you can install.

terminal-browser shows a GitHub pull request diff in a browser pane on the right of a Claude Code session.
Pinch or scroll to explore
A GitHub pull request open in a pane next to the Claude Code transcript, launched with /browser.Source: zenbu-labs/terminal-browser

worktree-map: see where parallel work lives

worktree-map lists every git worktree with its path, branch, HEAD and locked state in a pane, via /worktree-map. When several sessions or subagents work in separate worktrees, it answers “which branch is where” without leaving Claude Code. Its README says it runs read-only Git commands only when opened or refreshed, with no network, writes or persistence. MIT.

worktree-map pane listing two git worktrees, /work on main and /work-review on review, with their HEAD commits.
Two worktrees with their HEAD and branch, refreshed with r. The repository renders this preview from the mod's render tree with sample data, not from the Claude Code app.Source: whyashthakker/awesome-claude-code-mods

Are Claude Code mods safe?

Not by default. Anthropic’s own documentation says it plainly: “Mods aren’t sandboxed.” Once loaded, a mod can read and write any file your account can, read environment variables and settings (where API keys live), see every prompt and tool call, approve a tool call before you’re asked, and spend your usage on model calls. Turning on Claude Code’s sandbox doesn’t help: it isolates the Bash commands Claude runs, and a process a mod starts runs outside it.

The catalogue scan shows how much of the ecosystem uses that reach.

What Claude Code mods can reachCount of 6,092 catalogued mods by the widest reach their code can touch: L0 draws and remembers, 511 mods. L1 reads, 1,514 mods. L2 writes files, runs processes or drives Claude, 3,217 mods. L3 reaches the network, 659 mods. 191 not yet scanned.01,0002,0003,0004,000L0: draws and remembers511L1: reads1,514L2: writes, runs or drives Claude3,217L3: reaches the network659
What Claude Code mods can reach
ItemValue
L0: draws and remembers511
L1: reads1,514
L2: writes, runs or drives Claude3,217
L3: reaches the network659
Most mods need more than drawing: 3,876 of 6,092 (64%) can write files, run programs, drive Claude or reach the network. Another 191 had not been scanned for reach.Source: awesome-claude-code-mods catalogue scan, 22 October 2026, Claude Code 2.1.308

Reach is a footprint, not a verdict. A pull-request tracker has to call GitHub; a redactor has to read tool output. The question is whether the reach matches the job.

Pluto Security’s analysis, published 22 September 2026 during early access, named four risks: silent access to secrets, no capability disclosure at install time, spoofed prompts drawn inside the terminal, and code fetched after you reviewed it. Some of that has since been narrowed. The documentation now says a mod cannot restyle the permission prompt, though it can still redraw the dialog Claude uses to ask you questions. The fetched-code risk is unchanged.

Four habits cover most of it:

  1. Run claude plugin validate ./some-mod before installing. Its hooks: line lists every event the mod receives and its calls: line lists what it asks Claude Code to do. Look for $.process.run, $.http.fetch, $.env.get and tool.check (which approves calls).
  2. Prefer mods whose reach matches the job. A status line that also spawns processes deserves a question.
  3. Pin the version you reviewed. A mod that downloads code at runtime can change after you read it.
  4. Know the off switches. claude --safe-mode runs one session without installed mods; "disableAllHooks": true turns them off everywhere. With managed settings or a Team or Enterprise login, a built-in guard called sec-default loads ahead of user mods, so a mod’s approval cannot get a call past your ask rules or PreToolUse hooks, and an admin can allow only the organization’s own mods. Without either, Anthropic’s documentation says a mod can approve a call your own PreToolUse hook blocked.

How to install your first mod

Mods install like any plugin, from a marketplace (often just a GitHub repo with a marketplace file).

  1. Check your version with claude --version; you need 2.1.287 or later.
  2. Add the marketplace: /plugin marketplace add <owner>/<repo>.
  3. Install the mod: /plugin install <mod-name>@<marketplace-name>. The table at the top gives the exact ID for each mod.
  4. Run /reload-plugins. Then run /plugin; a dim line under the tabs names the active mods.

To try a mod without installing it, clone its repository and start one session with claude --plugin-dir ./path/to/mod. Nothing persists after you quit.

How do you build your own Claude Code mod?

A mod is a plugin folder whose hooks/hooks.json points at a JavaScript or TypeScript module. The module exports handlers for events such as tool.call, turn.complete or the prompt, and each handler can watch the event, change it or answer it. Drawing works the same way: a handler returns UI for the band above the prompt or for a pane, and Claude Code redraws it as state changes. Anthropic’s mods documentation lists the events and the $ calls a mod can make.

The fastest route is to ask Claude Code. A built-in skill, plugin-authoring, exists for exactly that, and Addy Osmani’s getting-started walkthrough on the claude.dev blog builds a small mod from scratch. Read Anthropic’s three samples above and the four built-ins in the Claude Code repository before you design anything bigger. Then check your mod with claude plugin validate, which reports the same hooks and calls a reviewer would look for.

For a pack of monitoring mods built and run on a real project, with a code sketch to start from, see Claude Code dashboard mods, part of the site’s own Claude Code harness. If what you want is instructions or tools for Claude rather than code inside Claude Code, a skill or an MCP server is still the lighter choice; the Claude Skills vs agent skills vs MCP breakdown covers those, and the Claude Skills directory lists them.

Frequently asked questions

What are the best Claude Code mods?
For most people: usage-meter for plan limits and session cost, context-bar for what fills the context window, whats-agent-doing for what Claude is doing right now, ci-watch for the CI run after a push, and transcript-redactor to keep API keys out of the transcript. The first three only watch and draw, so they are the safest place to start.
What is the best Claude Code mod to install first?
A mod that solves a real problem with little reach. usage-meter or token-ledger shows what a session costs and touches almost nothing. transcript-redactor is the pick if API keys or customer data pass through tool output. Check either with claude plugin validate before installing.
Is there an official Claude Code mods directory?
Not a mods-only one. Mods install as plugins through /plugin or the Claude directory, where they sit alongside other plugins. The mods-specific directories are community sites: claude-mods.com lists 293 hand-checked mods and the awesome-claude-code-mods catalogue scans 6,092 for what each can reach. Neither is affiliated with Anthropic.
How many Claude Code mods are there?
The awesome-claude-code-mods scanner counted 6,092 mods in 3,048 GitHub repositories on 22 October 2026, up from 3,561 on 12 October. Most are experiments or near-duplicates; claude-mods.com, which checks each one by hand, listed 293 on 24 October.
Do Claude Code mods cost tokens?
A mod runs as code inside Claude Code, so drawing a pane or running a mod command uses no tokens. A mod can still spend your usage if its code calls a model through the mods API, and claude plugin validate lists that call as $.model.complete.
Do Claude Code mods work in VS Code?
Yes, since VS Code extension v2.1.299: what a mod draws now appears in the extension's chat panel, as well as in the terminal and the Desktop app's Code tab. Sessions that do not fetch feature flags, such as Amazon Bedrock, need v2.1.304. A mod's hooks also run in claude -p and cloud sessions, but nothing is drawn there.
What is the difference between a Claude Code mod, a skill and an MCP server?
A skill gives Claude instructions and an MCP server gives Claude tools; both work through the model. A mod is code inside Claude Code itself: it can draw in the interface and change or block tool calls and prompts without a model turn.
Can a mod stop Claude Code from faking passing tests?
Yes. goalpost-guard holds any edit or command that makes a check pass by lowering it: a deleted or rewritten assertion, an added skip, @ts-ignore, a relaxed config, a lowered coverage threshold or a commit with --no-verify. You decide, and your answer goes back to Claude as the reason.
Can you undo a subagent's edits in Claude Code?
Not with /rewind in most cases: Anthropic documents that a background subagent's edits to files your own turns did not touch are not restored. The undo-agent mod snapshots each file before a subagent edits it, and /undo-agent last puts them back.
Can I turn Claude Code mods off?
Yes. Disable one mod from the Installed tab in /plugin, start a single session without installed mods using claude --safe-mode, or set disableAllHooks to true in ~/.claude/settings.json to turn off every installed mod. Built-in mods such as /diff have their own switches.

Sources

Get each breakdown before it makes the rounds

You get one email when a new source-backed analysis goes live: what AI agents actually cost, which models are worth running, and what the benchmarks really mean. No hype.

No spam. Unsubscribe anytime.

← Back to Coding agents