Command center for your Claude Code fleet.
You're running 8 agents across tmux. Three are blocked on permissions. One is asking a question. Two compacted 10 minutes ago and you didn't notice. You've burned $47 today and the rate is climbing.
Alt-tab. Scroll. Alt-tab. Scroll. Alt-tab. Approve. Wrong pane. Alt-tab.
clorch fixes this. One dashboard. Every agent. Every permission. Every dollar.
brew install lazypower/tap/clorchOr with Go:
go install github.com/lazypower/clorch/cmd/clorch@latestOr build from source:
git clone https://github.com/lazypower/clorch
cd clorch
make build # → bin/clorchclorch initThat's it. This installs async hooks into ~/.claude/settings.json and drops two bash scripts into ~/.local/share/clorch/hooks/. Your existing hooks are preserved — clorch merges non-destructively and backs up your settings first.
Start your agents however you normally do. tmux, iTerm tabs, whatever. clorch discovers them automatically through hook events.
clorch # launch the dashboardClaude Code ──hook events──▶ bash scripts ──atomic JSON write──▶ /tmp/clorch/state/
│
clorch TUI ◀──fsnotify────────────────────────────────────────────────────┘
No API calls. No terminal scraping. No process injection. Claude Code's native hook system fires events, two bash scripts write JSON state files, and clorch watches the directory. Sub-100ms updates on both macOS and Linux.
Real-time visualization of how full each agent's context window is. An 8-character bar gauge in the agent list and a 16-character gauge in the detail panel, color-coded green/yellow/red at 60%/80% thresholds. Shows compact count alongside the gauge so you can see context pressure at a glance. Opus gets 1M capacity, Sonnet/Haiku get 200K.
Spawn worktree branches with b and clorch renders the parent→child relationship as a tree in the agent list. Label any agent with l for quick identification.
Press d to open the detail panel. Every tool call, permission request, prompt, subagent spawn, and compaction is logged with timestamps. Newest events at the top, constrained to fit the viewport.
Individual subagents are tracked with their type, status, and ID — not just a counter. The agent list shows running subagent count, and the detail panel lists each subagent with a status indicator. Hook scripts sanitize agent IDs and maintain a full subagent map.
Each agent shows its running cost computed from Claude Code's JSONL transcripts. Input/output/cached token breakdowns in the detail panel. Aggregate burn rate in the header.
The agent table adapts to terminal width. Core columns (indicator, name, status, stale duration) always show. Optional columns — sparkline, branch, context gauge, cost, tool count, subagent indicator — appear progressively as width allows.
Agents approaching context limits get visual warnings — yellow at 3 compactions, red at 5. Helps you notice when an agent is churning through context before it falls off a cliff.
Press W to rename the tmux window for the selected agent. Pre-filled with the current window name. All tmux operations validate targets against injection characters.
The detail panel tracks every file modified by Edit/Write tools during a session. Most recent files shown, paths shortened relative to the agent's working directory.
When you upgrade clorch, the footer warns you if installed hooks are from an older version. Run clorch init to update them.
| Key | Action |
|---|---|
j / k |
Navigate agent list |
Enter / → |
Jump to agent's tmux pane |
a–z |
Focus action item |
y / n |
Approve / deny focused permission |
Y |
Approve all pending permissions |
! |
Toggle YOLO mode |
s |
Toggle sound notifications |
d |
Agent detail panel |
l |
Label selected agent |
b |
Branch selected agent (worktree) |
i |
Inject prompt into agent |
W |
Rename tmux window |
? |
Help |
q |
Quit |
~/.config/clorch/rules.yaml:
yolo: false
rules:
# Always approve read-only tools
- tools: [Read, Glob, Grep]
action: approve
# Never auto-approve destructive commands
- tools: [Bash]
pattern: "rm -rf"
action: deny
- tools: [Bash]
pattern: "git push --force"
action: deny
# Auto-approve edits
- tools: [Edit, Write]
action: approveFirst match wins. Deny rules always require manual review, even in YOLO mode.
clorch # TUI dashboard (default)
clorch init # install hooks
clorch init --dry-run # preview hook changes
clorch uninstall # remove hooks
clorch status # "3 working, 1 waiting" (for scripting)
clorch list # non-interactive agent table
clorch tmux-widget # status-right output with Nord colors
clorch version # print versionAdd to your ~/.tmux.conf:
set -g status-right '#(clorch tmux-widget)'
Gives you:
●3 ◉1 ✕0
Three working, one waiting, zero errors. Nord palette colors.
When an agent needs attention, clorch fires:
- Terminal bell — works everywhere
- macOS notification — native
osascriptbanner - System sound — Sosumi for permissions, Ping for questions, Basso for errors
Notifications only fire on state transitions, not repeated polls. Toggle sound with s in the TUI.
clorch parses Claude Code's JSONL session transcripts and shows:
- Total cost for today's sessions
- Rolling 10-minute burn rate ($/hour)
- Per-model token breakdowns
- Per-session cost in the agent table and detail panel
Displayed in the TUI header. No API keys needed — it reads the local transcript files directly.
| Variable | Default | Purpose |
|---|---|---|
CLORCH_STATE_DIR |
/tmp/clorch/state |
State file directory |
CLORCH_SESSION |
claude |
tmux session name |
CLORCH_TERMINAL |
auto-detect | Force terminal backend |
CLORCH_POLL_MS |
500 |
Polling fallback interval |
CLORCH_RULES |
~/.config/clorch/rules.yaml |
Rules file path |
- Go 1.21+
jq(used by hook scripts)- tmux (for approve/deny and pane navigation)
- macOS or Linux
Rewrite of androsovm/clorch (Python/Textual). Same state protocol, same hook scripts, same rules format. Drop-in replacement — no hook reinstall needed if you're migrating. As of v0.7.0, includes ported features from upstream through v0.4.0 (context gauge, subagent hierarchy, responsive columns, tmux safety) adapted to Go/BubbleTea architecture.
MIT