A mod is a Claude Code plugin made of function hooks. It can draw a live pane beside the
transcript, a band above the prompt, or a toast — and it can block, rewrite or react to any tool
call. Every one below was cloned and had its own test suite run here — 469 tests, 0 failures —
and the videos are rendered straight from the mods' own drawing code. Four of the fifteen are
also installed and enabled as plugins in the session that built this page; the rest are verified
by their suites and by running their drawing code, and are marked accordingly.
claude plugin install takes a -y flag that auto-accepts a command the
marketplace asks to run, without showing it to you. That prompt is the security
boundary. Install without it and read whatever surfaces. If you must accept one, use
--accept-command <sha256> with the hash the --json output reports.
Installed is not the same as loaded
Plugins load at session start only. An install mid-session does nothing until you restart.
Check what is actually live with claude plugin list — a ✔ enabled next
to the name means it will load next session.
All fifteen, tested here
burn-meter
Useful
A band above the prompt with a fire bar that grows with what the session has spent, your
5-hour and weekly plan-limit bars with reset times, and the total converted into burritos and
McDoubles. /burn opens the full panel with per-turn cost, and an alert fires each
time you cross a threshold.
Reads your session's own cost and rate-limit figures. No network, no processes, no files,
no model calls.
Hooks Bash before it runs and classifies the command: risky rm -rf,
git push --force, git reset --hard, DROP and
TRUNCATE through a live SQL client, supabase db reset,
vercel --prod, chmod -R 777, and downloaded scripts piped into a
shell. A match opens a red-alert pane with a one-time code — type it to arm, then press LAUNCH.
Anything else is denied and Claude is told why.
A speed bump, not a security boundary. It checks Bash commands only, and will
not stop an agent that writes a script and runs it another way. Keep your normal permissions on.
/launch-codes test <command> shows the verdict without running anything.
Redraws the whole transcript. Tool calls get line icons and a spinning ring while they run,
edits become diff cards with numbered changed lines, shell output lands in a terminal card with
stderr kept apart and long output folded. Seven skins, light and dark, plus a settings page and
a band above the prompt with context and plan limits.
Your own agent can design a new skin with you. /skin switches parts.
The context window as one row above the prompt, drawn the way Claude Code draws its own
meters. Filled cells first, the window left in dim cells, the reserve auto-compact keeps at the
end, then percent used, tokens over the window, and tokens left before auto-compact.
It follows /clear, resumes, branches, compaction and model switches, and re-reads
the engine every two seconds so a change nothing announces still shows.
Every agent step in plain English, with a running estimate of time saved. Files read as a
sentence ("Reading the pricing page to find the old numbers"), shell commands explained, test
results and durations spelled out. Built for training sessions where the audience has never
seen an agent work.
The default is rule-based and free. /narrate smart spends a model call per step,
sending the tool name and short fields — never file contents.
Keeps your subagents in a side pane beside the conversation: every agent a session runs, what
it is doing right now, what it costs in tokens, and its whole conversation one click away.
Running agents group under their parents, and finishing agents make one toast — by name when
alone, counted when several land at once.
/flow opens a live tree of the session's subagents beside the transcript. Each row
is one agent: status, type, name, elapsed time, what it is doing right now, its call count, and
its tokens once it finished. Agents waiting on your approval are highlighted, as are tool calls
over 30 seconds and agents silent for two minutes.
Where a pane cannot be drawn — a -p run, the VS Code extension — /flow
prints the same tree as text.
/swarm opens mission control for subagents and agent teams. An orchestration tree
nests every agent under whoever spawned it, with its type, model, live action and tool-call
count. Below it: a timeline of how the agents overlap, arcs and a log for messages between
teammates, and an activity feed. A status line keeps the live count, and a toast sums up the run
when the swarm stands down.
/race start <race> [name] in two or more sessions puts them on the same
track. The pane shows each session's tool calls, files touched and test runs live, and
/race done crosses the finish line. The honest way to compare two models or two
prompts on the same task.
Keeps its scores in ~/.claude/agent-race/. No network.
While Claude works, Clawd the crab acts out the spinner word above the spinner line.
All 189 of Claude Code's spinner words get their own scene — Baking slides bread into
an oven, Gitifying draws a commit graph — plus 1,890 lines of speech in a bubble that
follows him around, Rocky-from-Project Hail Mary style.
Costs nothing: every scene is drawn locally from code. No model calls, no tokens, no network.
/clawd-talk off shuts him up.
When a test run fails, a pixel boss spawns with one HP per failing test. Every later run that
fixes tests lands a hit. Zero failures is a KO with loot. It reads vitest, jest, pytest, mocha,
bun, cargo and go output, and sums each crate or package separately.
A pixel pet in a pane. It eats when Claude calls tools, gets sick when they fail, panics when a
destructive command shows up, falls asleep when the session goes idle, and evolves from egg to
crowned as you ship. Two days unfed and it is starving, hurt and sad.
/wrapped plays an animated stat reveal — session length, tool calls, MVP tool,
red-to-green test runs, longest turn, cost — then writes a 1200px PNG card to your Desktop.
Week and month totals come from your local transcripts.
Runs python3 to read its own transcripts. Nothing leaves the machine.
While the pane is open, a model call every so often turns the recent prompts and tool calls
into one dry line, typed out live. Harmless, and weirdly useful for noticing when the agent is
flailing.
Sends a summary of recent tool calls and the first 120 characters of each prompt to your own
Claude model, and only while the pane is open.
Play-by-play commentary on the session, spoken aloud with generated crowd audio — cheers when
tests pass, groans on fouls. /caster booth opens the booth pane.
Loud, and you have been warned. Sends a summary of recent tool calls to your own Claude model.
Not screen recordings. These frames are rendered by calling each mod's own drawing code —
frame() from clawd-spinner, arena() and step() from
boss-fight, fireCells() from burn-meter, petFrame() from code-pet — then
encoded with ffmpeg. What you see is what the mod puts on screen.
clawd-spinner — Clawd acts out Baking, Beaming and
Gitifying, each word with its own scene, with his speech bubble typing itself out beside him.
boss-fight — a vitest run fails with 7 tests, spawning a 7 HP boss. Later runs
land hits as failures drop 7 → 5 → 6 → 3 → 1, then a KO at zero.
burn-meter — the fire bar filling as a session spends its way from nothing to
nine dollars, flames growing taller as the heat rises. This is the band that sits above your prompt.
code-pet — one critter working through its moods: idle, eating, happy, level up,
sick, panic, asleep. Each mood is its own sprite and caption.
How these were tested
Each mod was cloned, its suite run on this machine with claude plugin test, and the raw
output kept. Every report below is the unedited test run. 469 tests, 0 failures.
Two different kinds of proof on this page
Suite-verified (all fifteen): the mod's own tests were run against this engine
version and the output is published above. That exercises the logic and the drawing code, and it
is what the videos call directly. Installed and running (four):context-view, agentpane, agent-flow and
launch-codes are also enabled as plugins in the session that built this page, so
they are proven to load and draw in a live session too. The other eleven are not installed here —
plugins load at session start, so installing them was a separate step this page does not claim.
One mod in the same collection did not make the cut
inbox-alerts fails 2 of its 8 tests against this engine version. It calls
prompt.submit from a command.run hook, which the engine rejects because
it would wait on the very turn the hook is holding. It is left off this page rather than shown as
working.
Read this before you install any mod
A mod runs inside your session
These are not sandboxed. A mod can read what the model reads, see every tool call, run
processes, and reach the network. Treat one like a dependency you are choosing to trust, not
like a theme.
Validation is not a safety check
The community catalogue of 1,018 mods runs claude plugin validate on each, which
reads the source without running it. It reports what a mod is capable of, not what it
intends, and it does not vouch for runtime behaviour.
Draw only — no network, no processes, no model calls
burn-meter
launch-codes
boss-fight
code-pet
clawd-spinner
swarm
Read-only, or writes only its own scores
context-view, agentpane, agent-flow — read session state and draw it
claude-skins — draws, and stores your skin preferences
agent-race — reads and writes ~/.claude/agent-race/
Runs a process, or calls your model
session-wrapped — runs python3 over your local transcripts, writes one PNG to your Desktop
inner-monologue, sportscaster — send a summary of recent tool calls to the same Claude model your session already uses
agent-narrator — only with /narrate smart: tool name and short fields, never file contents
Keep secrets out of commands
The mods that call a model send tool names, file paths and short command text. If you inline a
token or password in a shell command, that goes with it. Keep secrets in environment variables.
None of these fifteen makes its own network request, and none phones home.