A status line for AI coding agents in the terminal: the full version for Claude Code, a matching built-in preset for Codex. Two lines by default, one line or your own arrangement if you prefer, for dark or light terminals. It also shows how automodel routes each prompt when you use it.

_{Overview (2 min) · Documentation site}

• Line 1: where you are (directory, git branch and status, session name) on the left; who answers (output style, model, reasoning effort, automodel's routing) on the right.
• Line 2: budgets (context window, 5-hour and 7-day limits) on the left; session stats (prompt cache, cost and duration, lines edited) on the right. Both right blocks line up.

What it shows

Segment	Details
Directory	In Claude Code's own accent colour.
Git	Branch, ahead/behind, then staged, modified, untracked, conflicts, stash (see Glyphs). Worktrees get their own icon. Cached 2 s per directory.
Model	Claude Code's model, or, when automodel routes the session, the model it picked: jev → Opus 5.5.
Effort	5-step gauge coloured from low to max, dots ●●○○○ by default (the effort automodel picked, when it routes). Ultracode gets its own effect: by default Claude Code's violet with a sweeping highlight; else a drifting rainbow, or plain violet.
Route	Only when automodel routes the session: how it decided. The confidence 0.86 (green, yellow, red), or default, pinned, ⚠ fallback; ⚠ budget over automodel's spending cap; ⚠ jev: <why> when its router could not be asked; ↻ switched for a moment after it changes its mind. See the route part.
Context	Gradient bar and percentage; the percentage pulses above 85 %.
5h / 7d	Usage bar and percentage, the time until reset, plus its day and local time when it is not today (↻23h19 (Tue 9:00)). Then, in red, ⚠ full in 1h20 when at the pace so far the limit is reached before the reset. Once the reset time has passed, the window shows as reset (0 %, the 7-day one with its next reset) until Claude Code sends new figures.
Prompt cache	🔥 51m while warm, ⏳ 4m in the last sixth of its TTL, 🧊 cold once expired. A cold cache makes the next turn re-read the whole context.
Cost, duration, edits	From Claude Code's session counters.

Responsive: every segment has progressively shorter forms. When the terminal narrows, each line gives up the least useful detail first (session name, edits, routing details, labels), then bars turn into compact gauges, then the branch name is truncated. It stays readable down to about 60 columns, and the width comes from the terminal Claude Code runs in.

Animation: Claude Code re-renders a status line at most once per second (refreshInterval cannot go below 1), so the ultracode effects are designed for one frame per second: a fine gradient that drifts a little each second instead of jumping.

Fast: one jq call per render and cached git and terminal lookups: a few tens of milliseconds, so the 1-second refresh the installer sets (needed for the animations) is cheap. automodel, when present, is asked while the rest renders.

Install

Linux, macOS (needs bash 5, jq, git and curl; on macOS first brew install bash jq):

curl -fsSL https://raw.githubusercontent.com/moukrea/agentline/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/moukrea/agentline/main/install.ps1 | iex

It finds Git Bash, which Claude Code on Windows needs anyway (or installs Git for Windows with winget), installs jq with winget if it is missing, then runs the same installer in Git Bash, with the defaults: change them with agentline configure from Git Bash. Installer options go in $env:AGENTLINE_ARGS (e.g. '--glyphs unicode'). See Requirements.

This installs the latest release, never unreleased code from main.

The first install opens a full-screen setup assistant. At the top, a live preview: the real status line, drawn from your last Claude Code session (a sample one on a fresh machine) in the directory you run it from, with its real git state (simulated outside a repository), at your terminal's exact width. The session is read once, so only the animations move. It re-renders on every change, every second (animations run) and when you resize the window, so you see the responsive layout at work. "Preview as" switches the preview to ultracode, near-the-limits, a cold prompt cache or a session routed by automodel; the preview also narrows itself while you pick compact gauges, shows ultracode while you pick its effect and automodel's routing while you are on its settings. Defaults are Nerd icons and capsule bars when a Nerd Font is installed, Unicode otherwise, with a link to the font guide. Below, with ↑↓ ←→ and space:

• glyphs, theme (dark or light background), layout (two lines, one line, or a custom one kept from your config), progress bars, compact gauges (narrow terminals), effort gauge, ultracode effect, branch and reset icons, automatic updates;
• which parts Claude Code shows (directory, git, session name, model, effort, automodel's routing details, context, limits, cache, cost, edits…);
• automodel routing on or off, with a line saying whether automodel was found;
• for Codex, "Mirror Claude Code" (on by default) ticks the Codex items closest to the parts shown on the Claude Code side; the item list folds out to pick by hand.

Enter saves and installs for Claude Code and Codex, whichever are on the machine.

Non-interactive (CI, dotfiles), with options passed through bash -s --:

curl -fsSL https://raw.githubusercontent.com/moukrea/agentline/main/install.sh \
  | bash -s -- --yes --glyphs nerd --bar capsule --theme dark --layout two

From a clone, ./install.sh does the same. ./install.sh --help (or agentline install --help) lists every option.

The installer is idempotent: run it again at any time, it only rewrites what differs. It backs up each file it edits once (settings.json.agentline-backup, config.toml.agentline-backup) and keeps any status line you had configured, so uninstalling puts it back.

Pin a version, or read it first

A given release:

curl -fsSL https://raw.githubusercontent.com/moukrea/agentline/v0.5.0/install.sh \
  | AGENTLINE_REF=v0.5.0 bash

Turn automatic updates off (in the assistant, or agentline install --auto-update off) to stay on it. AGENTLINE_REF works for agentline update --force too, to go back to a release.

To read the code before running it, clone the release and run its installer: it installs that copy and downloads nothing.

git clone --depth 1 --branch v0.5.0 https://github.com/moukrea/agentline
cd agentline
less install.sh claude/statusline.sh   # the installer, and what Claude Code runs
./install.sh

The agentline command

Installed in ~/.local/bin:

agentline configure              # run the assistant again
agentline install --theme light  # change options without the assistant (--help: all of them)
agentline update                 # install the latest release if newer
agentline uninstall              # restore the previous status lines
agentline uninstall --purge      # and delete ~/.config/agentline
agentline version

Updates

With automatic updates on (the assistant's default), the status line checks for a new release at most once a day, in the background: rendering never waits for it. Updates only ever install a published release, and download nothing when you already have the latest one. Your configuration is kept across updates. Turn them off with agentline install --auto-update off, or update by hand with agentline update. See CHANGELOG.md for what changed.

Requirements and compatibility

• Linux, macOS or Windows, with bash 5, jq, git and curl (and ps, stty for the terminal's width). macOS ships bash 3.2: brew install bash jq. The installer runs itself with the first bash 5 it finds (Homebrew, Linuxbrew, MacPorts, then your PATH) and writes that bash's absolute path in Claude Code's statusLine, so an older bash first in Claude Code's PATH is never used. Windows: Git Bash (Git for Windows, bash 5) and jq, which install.ps1 installs; the status line runs Git's bash.exe by its C:/ path, and the terminal's width comes from the console (PowerShell, in the background, cached 5 s; 120 columns until the first answer). CI runs the full suite on Linux and macOS, and an install-and-render test on Windows. Python 3.11, when present, checks that an edited Codex config.toml still parses.
• Claude Code: agentline reads these fields of the JSON Claude Code sends to status lines: cwd and workspace (directory, project, worktree), model, effort.level, session_id, session_name, transcript_path, output_style, agent, vim.mode, fast_mode, context_window, cost, rate_limits and prompt_cache. A field your Claude Code version or plan does not send just hides its part.
• Codex: Codex draws its own line from tui.status_line in ~/.codex/config.toml; agentline only picks the items (see Codex).
• CLAUDE_CONFIG_DIR, CODEX_HOME, XDG_CONFIG_HOME and XDG_DATA_HOME are honoured. The agentline command goes to ~/.local/bin (AGENTLINE_BIN_DIR).

Fonts

Recommended: JuliaMono, with Symbols Nerd Font Mono as fallback. JuliaMono is the only free monospace font that draws every symbol Claude Code uses; the Nerd symbols provide the icons. Both are free for any use (SIL OFL 1.1, MIT).

The installer uses Nerd icons and capsule bars when it finds a Nerd Font on the machine, and Unicode glyphs otherwise. Installation per system and terminal settings: docs/fonts.md.

Configuration

~/.config/agentline/config, written by the installer and read on every render. Environment variables with the same names override it. Every setting also has an installer option (agentline install --help).

The three gauges (bars, compact gauges, effort) share one catalogue of styles: capsule (rounded ends, Nerd), smooth (eighth-cell bar on a solid rail), blocks █▒░, line ━╸─, segments ■□, braille ⣿⣀, ramp ▁▂▄▆█, dots ●○, bars ▰▱, squares ■□, pie (a single circle slice) and none (value only).

Variable	Values	Default
AGENTLINE_GLYPHS	nerd, unicode	nerd when a Nerd Font is installed, else unicode
AGENTLINE_THEME	dark, light: the terminal's background. light uses dark text and labels, a lighter bar track and darker greens, yellows and blues.	dark
AGENTLINE_LAYOUT	two, one, or a custom spec (see Layouts)	two
AGENTLINE_BAR	context, 5h and 7d bars: any style	capsule with Nerd glyphs, else smooth
AGENTLINE_COMPACT_STYLE	gauges when the terminal is too narrow for bars: any style	pie
AGENTLINE_EFFORT_STYLE	effort gauge: any style, or auto (same as the bars)	dots
AGENTLINE_ULTRA_EFFECT	rainbow (drifting gradient), violet (Claude Code's ultracode violet with a sweeping highlight), plain (static violet); applies to the effort part only	violet
AGENTLINE_BRANCH_ICON	auto, unicode ⎇, octicon, powerline, devicon	auto
AGENTLINE_RESET_ICON	auto, unicode ↻, octicon (history), mdi-history, mdi-progress-clock, mdi-refresh	auto
AGENTLINE_SEGMENTS	shown parts: dir git session meta model effort route ctx 5h 7d cache cost lines	all
AGENTLINE_AUTOMODEL	auto: show automodel's routing when it routes the session; off: never ask automodel (see automodel)	auto
AGENTLINE_CODEX_MIRROR	1: Codex shows the items closest to the Claude Code parts shown; 0: AGENTLINE_CODEX_ITEMS	1
AGENTLINE_CODEX_ITEMS	Codex items, in order (see codex/preset)	the preset
AGENTLINE_PATH_COLOR	R;G;B	215;119;87 (Claude Code's accent)
AGENTLINE_ICON_GAP	auto, 0, 1: space after Nerd icons, which are often drawn wider than their cell	auto (on with Nerd glyphs)
AGENTLINE_AUTO_UPDATE	1, 0	1

auto icons follow AGENTLINE_GLYPHS: octicons with nerd, Unicode otherwise. capsule needs Nerd glyphs; with unicode it falls back to smooth. AGENTLINE_CONFIG_VERSION is the installer's own bookkeeping: leave it as is.

Layouts

AGENTLINE_LAYOUT places the parts on lines. Two presets:

Layout	Lines
two (default)	dir git session | meta model route then ctx 5h 7d | cache cost lines
one	dir git | model route ctx 5h 7d cache, for a single line of status

Anything else is a custom spec: lines separated by ;, each one left | right (the right part is optional and is aligned with the right parts of the other lines). Part names are those of AGENTLINE_SEGMENTS; effort always goes with model, and unknown names are ignored.

agentline install --layout "dir git | model route; ctx 5h 7d | cost"
# or in ~/.config/agentline/config:
AGENTLINE_LAYOUT="dir git | model route; ctx 5h 7d | cost"

A part must also be in AGENTLINE_SEGMENTS to show. Whatever the layout, a line too wide for the terminal gives up detail in the same order, least useful first (session name, edits, routing details, labels…), each line on its own. The assistant keeps a custom layout from your config as "custom (from config)".

Glyphs

Git state	unicode	nerd (octicons)
staged	+	diff-added
modified, not staged	!	diff-modified
untracked	?	question
conflicts	=	alert
stash	$	stack
ahead / behind	↑ ↓	arrow-up / arrow-down
clean	✓	check

The Unicode set follows Starship's conventions.

automodel

automodel lets Claude Code pick the model and effort of each prompt: you choose Jev (auto) in /model, and it routes every prompt to, say, Opus at xhigh or Haiku at low. Claude Code itself only knows the session runs on "Jev (auto)" and reports its own effort, so a plain status line shows the wrong model and effort. agentline asks automodel instead:

~/projects/billing  ⎇ main ✓                                   jev → Opus 5.5 ●●●●○ xhigh  0.86 ↻ switched

• The model part shows jev → Opus 5.5: automodel's alias, then the model it picked for the last prompt, with the effort it picked in the effort gauge. When it routes to ultracode, the ultracode effect runs as usual.
• The route part (next to the model) shows how it decided, described below.
• Sessions on a model you picked yourself look as usual: no arrow, no route.

How it is wired

agentline owns Claude Code's statusLine; automodel only provides data. On each render, agentline finds automodel through the UserPromptSubmit hook automodel installs in settings.json (<automodel> --config <config.toml> hook decide), and runs automodel --config <config.toml> statusline --json with the payload Claude Code sent. The call runs while the rest of the line renders and is waited for 0.8 s at most; a missing, slow or broken answer just means "not routed": no error, no broken line. Where automodel lives, and whether it is recent enough, is cached until settings.json or the automodel binary changes, so without automodel a render costs no extra process.

automodel releases before 0.7.0 (without statusline --json) are detected (from automodel help) and never called: update automodel to see its routing.

Install order

Either order works:

• automodel first, then agentline: automodel's installer set its own status line; agentline's takes it over (and keeps it aside for --uninstall) and says that agentline now shows automodel's routing itself.
• agentline first, then automodel: automodel's installer sees a status line whose command contains agentline and leaves it in place (automodel doctor reports "agentline shows the automodel segment"); its updates leave it alone too.

Older automodel releases put their own status line back when they update: run agentline install again, and update automodel.

Uninstalling agentline while automodel is still installed gives automodel its own status line back (restored, or recreated from its hook), so you keep seeing the routing. If you had a status line of your own before agentline, the uninstaller prints the statusline_command = "…" line to add to automodel's config.toml so that automodel shows it before its segment. Uninstalling automodel leaves agentline in place; it simply stops showing routing.

With jaunt's rich view on, statusLine runs jaunt's wrapper, which runs agentline behind it: the installer reads which status line the wrapper chains and leaves it in place when it is agentline's.

If automodel's own status line runs agentline as its chained statusline_command (it sets AUTOMODEL_CHAINED=1), agentline does not ask automodel: automodel prints its segment itself in that setup.

The route part

Shown	Meaning
extend 0.86	Why automodel made its last decision, and how sure Jev, its router, is of it: green from 0.80, yellow from 0.60, red below. The why is how the prompt relates to the last one (continue, extend, inform, side question, aside, wrap up, new task, resume), with its probability, or asked (an effort asked in words), go-ahead, kept… with no number, or new with Jev's confidence in the effort. Needs automodel 0.19.2 or later; older ones show that confidence.
xhigh→high	The effort the last decision left, when it changed it.
default	Before the first decision of the session: automodel's default model and effort.
pinned	Your /effort wins: routing pauses until you set it back to its default.
⚠ fallback	Jev could not be asked; automodel used its default.
⚠ catalog	automodel cannot load its model catalog (the model part keeps Claude Code's).
xhigh	First, in red: the effort Claude Code displays (its spinner, /effort; ultracode in that mode), struck through, when automodel runs another one this turn (the one in the model part). Per-turn effort over a fixed base keeps the prompt cache. Needs automodel 0.9.0 or later.
⚠ budget	The session or the day is over automodel's spending cap (yellow).
⚠ jev: <why>	Why Jev could not be asked, such as a missing OpenRouter key.
↻ switched, ↻ compact, ↻ cold	For a moment (automodel's statusline_flash, 30 s) after a new decision: a model switch, a compaction, a cold prompt cache.

On narrower terminals the route part drops the effort left, then the why, keeping the number or word with a single ⚠ and ↻, then to a lone ⚠ when something is wrong, then disappears. Hide it altogether by leaving route out of AGENTLINE_SEGMENTS; the model and effort still follow automodel. AGENTLINE_AUTOMODEL=off stops asking automodel: the session then shows as Claude Code reports it.

AGENTLINE_AUTOMODEL_JSON replaces the call with a fixed answer, which is how the assistant previews and the tests show routing without running automodel; '{"v":1,"routed":false}' shows any session as not routed.

Codex

Codex has no external status line command: its tui.status_line setting is a list of items Codex draws itself, so custom glyphs, bars, colours, themes and layouts are not possible there. What you can choose is which items it shows. By default it mirrors the Claude Code side: model and reasoning, directory, git branch, session title, context used, 5-hour and weekly limits and estimated cost, minus whatever you hide in Claude Code (Codex has no counterpart for the route part). Turn mirroring off to pick items by hand (AGENTLINE_CODEX_ITEMS). status_line_use_colors = true is set as well.

The installer only touches the two keys it manages in [tui] (tagged # agentline). A status_line of yours is commented out, not deleted, and --uninstall puts it back.

How ultracode is detected

Claude Code reports ultracode to status lines as effort xhigh. The real signal is the ultra_effort_enter / ultra_effort_exit entry in the session transcript, or "ultracode": true in the settings. The transcript is read incrementally: only the bytes added since the previous render. In a session automodel routes, its answer says when it chose ultracode.

Privacy and network

agentline sends nothing about you or your sessions anywhere: no telemetry. On each render it reads, locally:

• the JSON Claude Code pipes to it (model, session, cost, limits…);
• the session transcript, only to detect ultracode, and only the bytes added since the previous render; the ultracode key of Claude Code's settings.json files;
• the git status of the session's directory, and the terminal's size;
• its config, and, when automodel routes the session, the answer of automodel statusline --json: a local process (what automodel itself does over the network is up to automodel).

It caches git state, the terminal and the last session (for the setup assistant's preview) in a private directory of yours: $XDG_RUNTIME_DIR/agentline-<uid>, else under $TMPDIR or /tmp.

The only network access is installing and updating. With automatic updates on, at most once a day the status line asks github.com for the latest release's tag (the API at api.github.com if that fails) and, only when there is a new one, downloads it from codeload.github.com. agentline update and the installer do the same; agentline install --auto-update off stops the daily check.

Troubleshooting

• Nothing shows up. Run the command Claude Code runs, with an empty session: eval "$(jq -r .statusLine.command ~/.claude/settings.json)" <<< '{}' prints a short line, or an error naming what is missing. jq and git must be in the PATH Claude Code starts with (started from an app or IDE on macOS, it may lack Homebrew's /opt/homebrew/bin). With CLAUDE_CONFIG_DIR set, run the installer with the same value. Then restart Claude Code.
• Wrong width. The line fits the terminal Claude Code runs in, found through its controlling terminal (ps, then stty size), minus Claude Code's padding. COLUMNS in Claude Code's environment wins; without it and without a terminal (some IDE panes), 120 columns are assumed.
• Boxes or question marks instead of icons. Your terminal font has no Nerd icons: install one (docs/fonts.md), or switch to agentline install --glyphs unicode. 🔥 ⏳ 🧊 come from your emoji font.
• Slow render. Time it in the project: time (eval "$(jq -r .statusLine.command ~/.claude/settings.json)" <<< '{}') takes a few tens of milliseconds once its caches are warm (run it twice). A huge repository slows git status (run at most every 2 s per directory: leave git out of --segments); a slow automodel is waited for 0.8 s at most (--automodel off). Releases before 0.5.0 were slow on bash 5.3: agentline update.

Development

tests/run.sh    # renders every fixture × glyph set × bar style at 17 widths, the
                # layouts, the light theme and automodel's answers (with a fake
                # automodel); installs, reinstalls, migrates, updates (from a fake
                # GitHub) and uninstalls in a throwaway HOME, including a simulated
                # `curl | bash` with scripted assistant answers and install over
                # automodel's status line

CI runs ShellCheck (install.sh, claude/statusline.sh, lib/wizard.sh, bin/agentline, tests/run.sh) and the suite on pushes to main, on tags and on pull requests.

docs/demo.gif is recorded from the real renderer with a pinned clock and sample session: tools/record-demo.py [--fonts DIR] [--emoji-font FILE] (JuliaMono and Symbols Nerd Font Mono files in DIR, ~/.local/share/fonts by default; needs Pillow, ffmpeg, jq and Noto Sans). Emoji need the CBDT (bitmap) build of Noto Color Emoji, which Pillow can draw, unlike the COLRv1 build many systems ship: get NotoColorEmoji.ttf from a noto-emoji release and pass it with --emoji-font. It never runs automodel: the routed scenes pass their answers in AGENTLINE_AUTOMODEL_JSON. Release tarballs leave out the demo, tools/, tests/ and .github/ (.gitattributes).

License

MIT © 2026 Emeric Commenge (moukrea).

The overview video's music: “Night Owl” by Broke For Free, CC BY 3.0.