Claude Code mods
English | Português (Brasil)
Your plan limits, your context and what the agent is doing, right above the prompt. Six
mods for Claude Code, installed from one marketplace.
limits-meter through a request, the /limits pane and a costly switch near the limit
(mp4). Drawn from the mod's real output by
scripts/screenshots/limits-meter-demo.sh.
Mods for everyday Claude Code UX. A mod is a Claude Code plugin made of function hooks: it can
draw a band above the prompt, a pane, a status line entry or a toast. Requires Claude Code
2.1.287 or later.
| Mod | Command | What it does |
|---|
| limits-meter | /limits | Plan limits, cache clock and context above the prompt: model and effort, 5-hour and weekly windows with reset times, a countdown of the prompt cache with an opt-in keep-warm that spends plan, context split by group (system, MCP, messages), tokens per turn, and a toast on a costly model switch near the limit |
| allowlist-coach | /allowlist | Counts permission dialogs per rule; after 5 approvals with no refusal, offers to add the rule to permissions.allow, asking before it writes |
| agent-watch | /watch | Subagents at a glance: a card per agent with its mascot, step and tokens, a band above the prompt, a toast when one stalls, and a summary naming the heaviest agent when they finish |
| test-hud | /test-hud | Test runs at a glance: passing over total on the hint line under the prompt, colored by the result, a sparkline of failures across runs, the failing tests grouped by file, and toasts when the suite turns green or a test turns flaky |
| launchpad | /pad | A menu of one-click actions under the header, above the prompt or in a pane: each button runs an installed command, skill or agent. ◆ pad opens a control panel: model and effort in one click, this collection's mods, the shortcuts. Pick and order up to 8 in /pad configuration, or ship a team's in the repository |
| plain-view | /plain-view | A quiet transcript: tool rows step aside and the agent's plan shows as a card above the prompt, with the current step, progress bars and a summary when the turn ends |
Using the mods
The images below are drawn from each mod's real output: scripts/screenshots/run.sh drives the
mod with sample data through claude plugin test and renders what it drew.
Every mod answers to the same four words, besides its own commands: /x on and /x off turn it
on and off in every conversation (the enabled option, the same as in /config; off keeps only
the command, so /x on can bring it back), and /x hide and /x show keep it on but take away
what it draws on its own (the band, the toasts, the status line, the card). The pane and the
command still open. /x help lists the rest. A mod's hide comes back in the next session.
limits-meter
The band shows on its own after the first response. Commands:
/limits open the pane: plan windows, context in tokens, the last 20 turns
/limits hide hide the band above the prompt, in later sessions too
/limits show bring the band back
/limits off | on turn the mod off or on in every conversation
/limits keep-warm on|off keep the prompt cache warm (asks first, spends plan)
/limits warn N warn on a costly switch with the 5-hour window from N% (default 70)
/limits help list the commands
The band leads with the model and effort; format: compact draws it as one line of numbers.
From 85% context the band and the pane show compact, which puts /compact [focus] in the
prompt and runs nothing. details on the band opens the pane. The pane also says when a window
reaches 100% at the current pace, before it resets, and draws the context fill turn by turn.
A switch to a bigger model (haiku < sonnet < opus < fable) or to effort max with the 5-hour
window at 70% or more shows one toast with the minutes left at the pace of the last 3 turns.
The cache cell counts down the prompt cache (⏱ 52:10), turns to the warning color in its last
10 minutes (the last minute of a 5-minute cache) and to ❄ cold once it lapses, with one toast.
It needs the cache TTL, which the engine gives on a model switch and which the mod otherwise
reads from a one-request turn after a pause; until it knows, the cell keeps the last turn's hit
rate. keep warm (or /limits keep-warm on) asks first, then sends one short message just
before the cache goes cold, a real turn drawn as one dim row. It stops at your next prompt or
after 2 hours. It spends plan: that is the only way this mod calls a model, and only when you
turn it on. The context bar is split by the engine's categories (system, MCP, messages), with a
legend under the band; the pane lists each category and each MCP server.
The cache clock and keep warm (mp4): the countdown after a reply, the context bar by group, the /limits pane, and keep warm pinging before the cache goes cold. Drawn from the mod's real output by scripts/screenshots/limits-meter-demo.sh.
Install
/plugin install limits-meter --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./limits-meter.
allowlist-coach
It counts on its own each time you answer a permission dialog. Commands:
/allowlist open the pane: every counted rule, numbered, ready ones first
/allowlist allow 1 add rule 1 to permissions.allow (asks first, and which file)
/allowlist dismiss 2 stop offering rule 2
/allowlist remove 1 take rule 1 back out of allow, when the coach added it (asks first)
/allowlist reset 2 set rule 2's count back to zero (asks first)
/allowlist reset clear this project's counts (asks first)
/allowlist demo the pane with sample rules; nothing is read or written (demo off exits)
/allowlist hide keep counting, with no toast or note in the dialog
/allowlist show bring them back
/allowlist off | on turn the mod off or on in every conversation
/allowlist help list the commands
The pane has tabs by status, a filter, and a scrolling list. A risky rule says why it is never
offered.
Install
/plugin install allowlist-coach --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./allowlist-coach.
agent-watch
The band above the prompt, the status line and the toasts show on their own while subagents run. Commands:
/watch open the pane: a card per agent with step, latest actions and tokens
/watch demo add three fake agents, one of them stalled, to see the pane
/watch clear drop finished and demo agents
/watch clear done drop finished agents only
/watch clear demo drop demo agents only
/watch hide keep tracking, with no toast or status line
/watch show bring them back
/watch off | on turn the mod off or on in every conversation
/watch help list the commands
Each agent has a card with its mascot and color, its step, the two calls before it, a bar of its
tokens against the heaviest agent's, and an estimate of its share of the 5-hour window and of the
week. The cards sit in two groups, Running and Finished; a bar across the pane folds its group. From
five agents (the compactAt option) each is one boxed line, and its name opens the card. A stalled agent has
an investigate button that asks about it in the prompt.
Install
/plugin install agent-watch --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./agent-watch.
test-hud
The tests line (atop the hint line under the prompt, red, yellow or green by the result) and the
toasts show on their own each time a test runner runs in Bash. Commands:
/test-hud open the pane: last run, failing tests (new ones marked), fixed tests,
sparkline, last 10 runs
/test-hud demo add six fake runs that go from red to green
/test-hud clear drop the runs
/test-hud hide keep counting, with no tests line or toast
/test-hud show bring them back
/test-hud off | on turn the mod off or on in every conversation
/test-hud help list the commands
In the pane, failing tests come grouped by file (a class for Maven and Gradle, a suite for jest and
bun, a module for cargo; go and rspec stay a flat list). The first group is open; press a header to
open or close one. show puts the file in the prompt (@src/a.test.ts), only where the group is a
real path (vitest, pytest); run asks for only that group's tests with the same runner. Press a
failing test to ask Claude for a fix, or run again to ask for the same command: all of them put
the request in the prompt and run nothing. Press a run in the list to see it; failing tests that
failed, passed and failed again are marked flaky?, and one toast names them, once per test
(flakyToast option, on by default). With the keepHistory option the runs carry over to the next
session in the same project.
Install
/plugin install test-hud --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./test-hud.
launchpad
The menu shows on its own when a session starts and after /clear, until the first prompt.
◆ pad, under the hint line below the prompt, opens the control panel: the session's
model and effort as buttons (a press runs /model or /effort; opus and max turn yellow
when the 5-hour window passes 70%), then two tabs: Mods, a switchboard with a toggle (/x on and
off) and an eye (/x hide and show) for each mod of this collection, and the shortcuts.
The labels follow the language option. Commands:
/pad show the menu again
/pad panel open the control panel, as ◆ pad does
/pad configuration pane to order, remove and add buttons, up to 8, and the mods to install
/pad list every button with what it runs
/pad add 🔎 Revisão | /code-review add a button for an installed command, skill or @agent
/pad remove 3 drop button 3
/pad reset back to the defaults
/pad place header | prompt | pane where the menu shows: under the header, right above the prompt, or in a pane
/pad hide | show hide the menu or show it again
/pad off | on turn the launchpad off or back on in every conversation
/pad quiet [on | off] quiet mode: no mod draws anything by itself; a chip on the hint line brings it back
Install
/plugin install launchpad --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./launchpad.
plain-view
Off until you turn it on (/plain-view on, the enabled option, or the switch in launchpad's
panel). While it is on, the transcript keeps the conversation and drops the tool rows that
worked; a failed or interrupted call, a permission dialog and the agent's questions always draw
in full. Above the prompt, a card follows the request: its title, Step 2 of 4 with a bar, one
row per task of the agent's list with its own bar (Done, ~40% for the current step, Next,
Up next). The current step's percentage is an estimate from its tool calls, hence the ~.
When the turn ends the card turns green with the time it took and the files changed and read,
or grey on Esc; the next request starts a fresh one, and [-] folds it.
The agents the main turn hands work to show in one row of the card (◇ 1 agent running · code-review · 3m 12s, with agent-watch's /watch when it is installed, and agent-watch's band
message under it: └ Agents 0/3 ░░░░░░░░░░ · ▲ Pre-outline · Writing pre-outline.md), and with no task list
the card waits for them (Waiting for 1 agent). A turn with no list that called tools or agents
ends in a small green card (✓ Done, the time, the files and the agents); a plain conversation
leaves none.
Three options shape what you see, all in /config:
agentText, what stays of the agent's messages: final (default: the messages written before
a tool call step aside, the final answer stays), none, card (the answer's first sentence in the end card) or
all. Failures, permission dialogs and the agent's questions always show. The agent's
thinking is drawn outside any mod's reach; "showThinkingSummaries": false in
~/.claude/settings.json hides it everywhere, ctrl+o included.askForTasks (on by default): one line in the system prompt asks the model to keep a task
list for work of more than two steps, or a checklist in its reply when the session has no task
tool, so the card shows each step; a few tokens per request. When a checklist in the text stops
moving, the card says list not updated and the model gets a short reminder with a tool result.
Off, the card reads the list the agent keeps on its own, or a checklist in its answer.request, how much of a request of several lines the card's title shows: line (default: the
first line and +5 lines), joined, auto (the first line of prose) or all. The card's
show request button unfolds the whole request.
/plain-view the settings pane: every option in three tabs, one click each
/plain-view on | off show or hide the card and the tool rows
/plain-view hide | show keep the rows folded, without the card above the prompt
/plain-view demo a sample plan in the card for 12 seconds
/plain-view palette the 8 bar palettes with a sample, and a button to switch
/plain-view palette aurora switch to one by name
/plain-view help list the commands
The bars are a gradient of the palette option's colors (default claude) with a shine that
runs while the agent works. Each palette has a dark and a light set, picked by your theme.
animation: off keeps the bars still, in the theme's own colors.
Four short scenarios, each drawn by the mod itself:
A task list (mp4)
No task tool: a checklist in the answer (mp4)
Esc interrupts the turn (mp4)
A background agent (mp4)
Install
/plugin install plain-view --marketplace ice-lfernandes/claude-code-mods
Or for one session, from a clone: claude --plugin-dir ./plain-view.
Install
Each mod's section above has its own install line. In a Claude Code terminal session:
/plugin install limits-meter --marketplace ice-lfernandes/claude-code-mods
Answer y to add the marketplace, then pick a scope.
To try a mod for one session from a clone:
claude --plugin-dir ./limits-meter
What each mod reaches
A mod runs inside Claude Code with your permissions and is not sandboxed. Read it before you
load it, and run claude plugin validate <folder> to list every event it hooks and every call
it makes.
| Mod | Network | Runs processes | Files | Calls a model | Sends data anywhere |
|---|
| limits-meter | No | No | No (on and off write its own option through /config) | Only if you turn on keep-warm: one short message per cache lap, until your next prompt or 2 hours | No |
| allowlist-coach | No | No | Reads and writes .claude/settings.local.json, or .claude/settings.json when you pick it, after you confirm. on and off write its own option through /config | No | No |
| agent-watch | No | No | No (on and off write its own option through /config) | No | No |
| test-hud | No | No | Reads Bash's saved copy of an output too long to show whole. on and off write its own option through /config | No | No |
| launchpad | No | No | Reads .claude/launchpad.json and the agent files in .claude/agents/, in the project and in your home folder. on and off write its own option through /config | No | No |
| plain-view | No | No | No (writes its own options through /config when you run on, off or palette, or press one in its settings pane) | No | No |
Developing
claude plugin validate ./limits-meter
claude plugin test ./limits-meter
claude plugin validate ./allowlist-coach
claude plugin test ./allowlist-coach
claude plugin validate ./agent-watch
claude plugin test ./agent-watch
claude plugin validate ./test-hud
claude plugin test ./test-hud
claude plugin validate ./launchpad
claude plugin test ./launchpad
claude plugin validate ./plain-view
claude plugin test ./plain-view
Every mod ships tests, including a render test on the terminal and desktop surfaces.
Checks before a merge
GitHub Actions (.github/workflows/ci.yml) runs on every pull request and every push to main,
with no Claude login and no model call:
claude plugin validate and claude plugin test for each mod, on a pinned Claude Code version;scripts/check-shared.sh: the shared files are the same in every mod;scripts/check-manifests.py: every JSON file parses, the marketplace lists every mod, and each
entry's name and description match the mod's plugin.json;scripts/check-versions.sh <base>: a mod changed in the pull request raised its version, since
Claude Code updates an installed plugin only when the version moves;- the screenshots match what the mods draw now (
scripts/screenshots/run.sh, then no diff);
shellcheck on the scripts.
Run the same checks locally before a push:
./scripts/check-shared.sh
python3 scripts/check-manifests.py
./scripts/check-versions.sh origin/main
./scripts/screenshots/run.sh && git diff --stat screenshots
Shared helpers
A mod installs alone and cannot import another mod's code. The helpers every mod needs live in
hooks/ui.tsx, tested by tests/ui.test.ts, and both files are copies, the same in each mod:
language and icon style from the options, a prompt fill with its [blank] marked, the window of
a long list, a row of clickable verbs, and the number formats. Nothing in it takes $: the
engine follows $ only into functions of the file that uses it, so the calls on $ stay in
each mod's register.tsx. A mod that names tool calls also carries hooks/phrases.ts and
tests/phrases.test.ts: a tool call in a few words, in both languages (reading routes.ts,
searching "invoice"). Change one copy, copy it to the other mods, then check them:
./scripts/check-shared.sh
Pane layout
Every pane follows the same outline, after launchpad's /pad configuration:
- A dim hint on top: what the pane shows and where the figures come from.
- Sections, each under a bold heading.
- An empty state that names the next step (
/test-hud demo shows what this looks like).
- A footer: the command's verbs as clickable buttons (
/test-hud clear · demo · help), then a
close button.
A button that runs something from outside the mod, or undoes the person's data, puts the text in
the prompt instead of running it.
Shared bands
Several mods can draw on the same line: AbovePrompt (limits-meter, plain-view, and launchpad
under /pad place prompt) and PromptHint (launchpad's ◆ pad). Each ui.render hook there calls
next(e) and stacks its own row with what came back, never in its place, so every mod's row
shows:
AbovePrompt: the mod's row first, the rows of the mods beneath it after. The one exception
is launchpad, whose row goes last so it sits against the prompt box.PromptHint: the engine's hint first, the mod's row or button after.- A mod with nothing to show, or switched off, returns
next(e) as it is.
Otherwise the order between mods follows the order Claude Code loads them; a mod cannot set it.
After a change to what a mod draws, regenerate the images:
./scripts/screenshots/run.sh
Custom mods, skills and agents
Need a mod, skill or agent built for your own workflow or team? I take on custom work: hooks,
panes, status lines, slash commands, skills and subagents for Claude Code.
YouTube
My channel, Java + AI: youtube.com/@JavaPlusAI
License
MIT. See LICENSE.
