 ██████╗ ██╗      █████╗ ██╗   ██╗██████╗ ███████╗████████╗██╗   ██╗██╗
██╔════╝ ██║     ██╔══██╗██║   ██║██╔══██╗██╔════╝╚══██╔══╝██║   ██║██║
██║      ██║     ███████║██║   ██║██║  ██║█████╗     ██║   ██║   ██║██║
██║      ██║     ██╔══██║██║   ██║██║  ██║██╔══╝     ██║   ██║   ██║██║
╚██████╗ ███████╗██║  ██║╚██████╔╝██████╔╝███████╗   ██║   ╚██████╔╝██║
 ╚═════╝ ╚══════╝╚═╝  ╚═╝ ╚═════╝ ╚═════╝ ╚══════╝   ╚═╝    ╚═════╝ ╚═╝

Instrumentation for Claude Code sessions.

Context fill, cost, cache ratio, compaction distance, and live tool activity — parsed from the session transcript and drawn in your terminal.

Every terminal block on this page is real output, captured from the tools in this repo. Nothing here is a mockup.

version: 0.8.7
runtime: Python 3.13+, stdlib only
dependencies: none
platforms: macOS, Linux, WSL 2
license: MIT
source: github.com/slima4/claude-tui

brew install slima4/claude-tui/claude-tui

claudetui setup

No Homebrew? curl -sSL https://raw.githubusercontent.com/slima4/claude-tui/main/install.sh | bash — see install.

Statusline

claudetui mode full

Claude Code renders a status line under the prompt on every turn. This replaces it with a readout of what the session is actually costing you.

full mode · 148 columnsreal capture

 0011110 ▏▮▮▮▮▮▮▮▮▎▯▯▯▯▯▯▯▯▯▯┆▕ C 42% ⚡421.3k/1.0M ⋮ ETA 13 turns ⋮ Opus 5 ⋮ ▅▃▅▇▆▇▆▃▂▁█▆▃▂▅▄ ⋮ $12.47 ⋮ ⏱ 2h 10m ⋮ CMP 1x ⋮ 54% eff ⋮ #a1b2c3d4
 0100011 ▏▮▮▮▮▮▮▋▯▯▯▯▯▯▯▯▯▯▯▯▯▕ S 34% ⏱ 1h59m ⋮ app ⋮ ⎇ feature/auth +180 -60 ⋮ TRN 35 ⋮ FIL 2 ⋮ ERR 0 ⋮ CAC 99% ⋮ THK 8 ⋮ ~$0.36/turn
 1000101 ▏▮▮▮▮▮▮▮▮▮▮▮▮▮▮▏▯▯▯▯▯▕ W 71% ⏱ 71h59m
         read auth.ts → edit auth.ts → bash npm → edit session.ts ⋮auth.ts×1 session.ts×1

 Opus 5 ⋮ ▏▮▮▮▮▮▮▋▯▯▯▯▯▯▯▯┆▕ C 42% ⚡421.3k/1.0M ⋮ ▏▮▮▮▮▮▍▯▯▯▯▯▯▯▯▯▯▕ S 34% ⏱ 1h59m ⋮ ▏▮▮▮▮▮▮▮▮▮▮▮▎▯▯▯▯▕ W 71% ⏱ 71h59m

Four rendered rows: context bar, 5-hour plan bar, 7-day plan bar, and the live tool trace. Every field is documented below. Narrow screens show compact mode instead.

C 42% ⚡421.3k/1.0M: Context fill and token count against the window. The bar's colour tracks the compaction ceiling, not 100% — it turns warm as you approach the point where Claude Code compacts.
ETA 13 turns: Estimated turns before compaction, from a fixed 33k-token buffer model and your recent per-turn growth.
CMP 1x: Compactions so far this session.
54% eff: Useful context divided by total context built. Falls as compactions discard work you already paid to load.
▇▄↓▅▅▃▅▇▆: Per-turn output sparkline. ↓ marks a compaction boundary; older bars fade.
S 34% ⏱ 1h59m: 5-hour plan limit used, and the countdown to reset. Pro/Max only — Claude Code supplies it, so there's no extra API call.
W 71% ⏱ 71h59m: 7-day plan limit and its reset countdown.
⎇ feature/auth +180 -60: Git branch and uncommitted diff stat.
TRN · FIL · ERR: Turns, files touched, and tool calls that came back as errors.
CAC 99% · THK 8: Cache hit ratio and thinking-block count.
$12.47 · ~$0.36/turn: Session cost and cost per turn, taken from Claude Code's own accounting rather than re-derived.
read auth.ts → edit auth.ts: The current turn's tool trace, with per-file edit counts on the right.
0011110: Matrix-rain widget. Swappable: matrix, hex, bars, progress, or none.

compact mode · one linereal capture

 Opus 5 ⋮ ▏▮▮▮▮▮▮▮▮▎▯▯▯▯▯▯▯▯▯▯┆▕ C 42% ⚡421.3k/1.0M ⋮ ▏▮▮▮▮▮▮▋▯▯▯▯▯▯▯▯▯▯▯▯▯▕ S 34% ⏱ 1h59m ⋮ ▏▮▮▮▮▮▮▮▮▮▮▮▮▮▮▏▯▯▯▯▯▕ W 71% ⏱ 71h59m

Same data, one row — claudetui mode compact.

A third mode, claudetui mode custom, opens a curses configurator where every field above is an independent toggle, saved to ~/.claude/claudeui.json.

Monitor

claudetui monitor

A dashboard for a second terminal. It finds the active session, follows the transcript file, and redraws when it changes.

claudetui monitor · 84 columnsreal capture

0100101011001100001110100101011111100011110000011011001101111010100101010001100011
────────────────────────────────────────────────────────────────────────────────────
  ● claude-sonnet-5  │  a1b2c3d4  │  2h 09m 35s  │  ACTIVE

  CONTEXT
  ▏▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▏▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯▯┆▯▕ C 42% ⚡421.3k/1.0M

  ▂▂▂▂▄▂▇▃▇▇▆▃▇▂▃▂▇▄↓▅▅▃▅▇▆▇▆▃▂▁█▆▃▂▅▄
  Compactions: 1  │  Turns left: ~13  │  Since compact: 16
  Efficiency: 54%  │  Wasted: 646.4k/1.4M

  ACTIVITY
  2 files edited  │  +0 / -0 lines
  16 read  │  21 edit  │  8 grep  │  15 bash  │  12 write

  CURRENT
  4 tools  │  Edit:2  Read:1  Bash:1
  read auth.ts → edit auth.ts → bash npm → edit session.ts │ auth.ts×1 session.ts×1
  auth.ts(1r/1e) session.ts(0r/1e)
  0 errors  │  1 thinking

  SESSION
  35 turns  │  72 tools  │  Edit:21  Read:16  Bash:15
  auth.ts(10r/25e) session.ts(6r/8e)
  0 errors  │  8 thinking (23%)  │  99% cache
  17:49:45  read auth.ts
  17:49:45  edit auth.ts
  17:51:10  $ npm test -- auth
  17:51:10  read auth.ts
  17:52:35  read auth.ts
  17:52:35  edit auth.ts
  17:52:35  $ npm test
────────────────────────────────────────────────────────────────────────────────────
  stats  details  log  waste  export  sessions  config  info  ?help  quit

Context bar and sparkline, compaction count and distance, per-turn tool trace with read/edit counts per file, session totals, and a scrollable event log.

s stats
d details
l log
w waste chart
e export
o sessions
c config
i Claude status
? help
q quit

In the log viewer, f cycles the filter — errors, bash, edits, search, agents, skills, compactions — and a toggles live follow. Navigation is vim-style: Ctrl+F/Ctrl+B, g/G.

Context efficiency

claudetui chart

Compaction is the expensive part of a long session: the window fills, Claude Code summarises, and everything below the summary is re-read from scratch. This chart splits each stretch between compactions into what the tokens actually bought.

claudetui chart · also the monitor's w keyreal capture

  CONTEXT EFFICIENCY — HORIZONTAL

    ▒▒ system   ▓▓ summary   ██ useful   ░░ headroom

    Seg 1  ▒▒▒██████████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  1.0M
          65.6k system │ 293.0k useful │ 641.4k headroom │ → compacted
          ──── compact #1 ────
  → Seg 2  ▒▒▒▓████████████████████  421.3k
          65.6k system │ 5.0k summary │ 350.7k useful

Per-segment breakdown against the full window.

▒▒ system: System prompt and tool definitions — paid for on every request, never reclaimed.
▓▓ summary: Context carried across a compaction boundary as a summary.
██ useful: Conversation and tool output that did work.
░░ headroom: Window bought but never used before the segment ended. The number to shrink.

API sniffer

claudetui sniffer

A local proxy between Claude Code and the API. Set ANTHROPIC_BASE_URL at it — or use the claudetui sniff launcher, which detects the port — and every request prints as one line.

claudetui snifferoutput format

  #1   POST /v1/messages  opus-5  132.6k->459  $0.237  25632ms  496.1KB/12.5KB  100%c  [Tt]
  #2   POST /v1/messages  opus-5  133.1k->422  $0.239  13451ms  498.4KB/22.7KB  100%c  [U]  Agent
  #3   POST /v1/messages  haiku-4-5  15.4k->258  $0.016  2018ms  56.7KB/9.5KB  0%c  [tUUUU]  Read,Read,Read,Read  +agent.1
  #4   POST /v1/messages  opus-5  134.5k->211  $0.242  16162ms  503.3KB/11.6KB  99%c  [tU]  Read
  #5   POST /v1/messages  opus-5  49.8k->1.1k  $0.653  26245ms  181.9KB/17.0KB  42%c  [TtU]  Read  compaction

Per request: tokens in and out, cost, latency, bytes on the wire, cache ratio, content-block types, tool names, and the sub-agent that made the call. Sub-agents are detected by the presence of the Agent tool, since they share the parent's session ID.

[Tt] [U] [tUUUU]: Content blocks in the response — T thinking, t text, U tool use, S server tool, W web search result, M/m MCP.
100%c · 0%c: Cache hit ratio for the request. Red at zero — a cold prompt you're paying full price for.
+agent.1: First request from a newly detected sub-agent.
compaction: A compaction request, including the ones Claude Code makes without telling you.

Requests are logged as JSONL to ~/.claude/api-sniffer/, created 0600, with API keys redacted unless you pass --no-redact. Each instance writes a port file, so several projects can run their own proxy at once.

Also included

Session stats

claudetui stats

Post-session analytics over the transcripts in ~/.claude/projects/.

cost breakdown
token sparkline
tool usage
file activity
--days 7 multi-session summary

Session manager

claudetui sessions list

Browse, inspect, compare, resume, and export past sessions.

list
show
resume
diff two sessions
export

Hooks

installed by claudetui setup

Three hooks that put project history in front of Claude before it acts.

SessionStart — file activity hotspots
PreToolUse — high-churn warning before an edit
PostToolUse — reverse dependencies after an edit

Slash commands

/tui:session

In-session reports, generated by Claude from the current transcript.

/tui:session
/tui:cost
/tui:perf
/tui:context

Install

macOS · Linux · WSL 2

Installing puts the statusline, hooks, and slash commands into ~/.claude/settings.json and symlinks claudetui onto your PATH.

# Homebrew
brew install slima4/claude-tui/claude-tui
claudetui setup

# or the install script
curl -sSL https://raw.githubusercontent.com/slima4/claude-tui/main/install.sh | bash

# or from a clone
git clone https://github.com/slima4/claude-tui.git
./claude-tui/install.sh

# pin an isolated interpreter
CLAUDETUI_PYTHON="$(uv python find 3.13)" ./install.sh

# install somewhere other than ~/.claude-ui
INSTALL_DIR="$HOME/tools/claude-tui" ./install.sh

# uninstall
claudetui uninstall
brew uninstall claude-tui

# already ran brew uninstall first?
curl -sSL https://raw.githubusercontent.com/slima4/claude-tui/main/uninstall.sh | bash

Windows needs WSL 2 — the tools want a Unix-like environment.