Claude Code Context Bar

A bar for Claude Code that shows the current model, how full the context window is, and how long until your prompt cache lapses, with what the next request costs either way.

Opus 5.5  [██████░░░░░░░░░░░░░░░░░░]  256.9k / 1M (26.0%)  ⏱ 58:59  $0.051→$2.056

It comes in two forms:

	Mod (mod/)	Status line (claude-context-bar.py)
Cache readout	A live countdown, ticking every second	The time the cache lapses, ⏱ til 1:24pm
Needs	Claude Code with mods (early access)	Any Claude Code from v1.0.71, and Python 3

The status line can't count down because Claude Code only reruns it on conversation events, never on a timer. Use the mod if your Claude Code supports mods, and the status line otherwise.

What it shows

• Model, then a progress bar of the context window that shifts from green to red as it fills:

  Range	Color
  0–15%	Bright green
  15–28%	Lime
  28–42%	Yellow-green
  42–55%	Yellow
  55–65%	Gold
  65–75%	Orange
  75–87%	Orange-red
  87–100%	Red

• Tokens in context against the window size.

• ⏱, the prompt cache, followed by two prices. Green is what the next request costs while the cache is warm (a cache read). Red is what it costs once the cache has lapsed and must be written again. Once it lapses, only the red price is shown.

The gap between the two prices is why the timer matters. A cache write costs 1.25× the input rate on a 5-minute cache and 2× on a 1-hour cache, while a read costs 0.1× on most models and less on some (0.05× on Opus 5.5, 0.025× on Fable 5.1). Letting a large context lapse costs 12.5× to 80× more than sending in time. Every request renews the cache, including each tool call.

Mod

Install

Clone the repo, then add its mod/ folder to the env block of your user settings.json (~/.claude/settings.json):

"env": {
  "CLAUDE_CODE_PLUGIN_DIRS": "/path/to/claude-context-bar/mod"
}

New sessions load it. To try it in one session only, run claude --plugin-dir /path/to/claude-context-bar/mod instead.

Mods are early access, behind a rollout flag. If the bar never appears, Claude Code may not be able to fetch that flag. This happens on a third-party API provider or with telemetry turned off, and claude --debug then logs hooks modules are not turned on for installed plugins. Turn mods on in the same env block:

"CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1"

If you used the status line before, remove statusLine from settings.json so the two don't draw side by side. Keep subagentStatusLine; the mod can't draw the subagent rows.

To update, git pull. New sessions load the new version.

Tested on Claude Code 2.1.284 and 2.1.287.

Options

Option	Default	Meaning
cacheTtl	1h	5m or 1h, matching how long Claude Code caches for you: 1 hour on a Claude subscription, 5 minutes on an API key, unless Claude Code's promptCacheTtl setting says otherwise. It also sets the write price: 2× input for 1h, 1.25× for 5m.
placement	below	below draws under the prompt, above the hint line. above draws over the prompt.
barWidth	24	Progress bar width in cells.

Change them in /config, or under pluginConfigs in settings.json:

"pluginConfigs": {
  "context-bar": { "options": { "cacheTtl": "5m" } }
}

What the countdown shows

• After each request it starts again from the full TTL, counted from when the request was sent. Subagent requests don't reset it, because they don't keep the main conversation's cache warm.
• At zero it reads ⏱ lapsed.
• When you resume a conversation (claude -c, --resume or /resume), the bar comes back with the conversation's size and how much time its cache has left, read from the transcript. You can see whether it is still cached before you send anything.
• After /compact or auto-compaction, it shows ~ before the new size, an estimate, and ⏱ compacted with the write price, since the summary isn't cached yet. The next request replaces both with the real figures.
• After /clear the bar is hidden until the new conversation's first request.
• ⏱ --:-- means the bar has a size but no time, because it couldn't read the transcript. This can happen in the desktop app with a very long session, over about 4 MB of transcript. The next request fixes it.

Status line

Install

git clone https://github.com/k-wolfe99/claude-context-bar.git
cd claude-context-bar
./install.sh

Then restart Claude Code. install.sh copies claude-context-bar.py and claude-agent-rows.py to ~/.claude/ and points statusLine and subagentStatusLine at them.

To update, run git pull && ./install.sh. No restart is needed, because Claude Code reruns the script on every repaint.

If statusLine or subagentStatusLine already points at a different command, the installer leaves it alone and says so. --force repoints it. Exit codes: 0 installed or updated, 2 couldn't read or write settings.json (the scripts are still copied), 3 left an existing entry alone.

To install by hand, copy the two scripts to ~/.claude/ and add:

"statusLine": {
  "type": "command",
  "command": "python3 /YOUR_HOME/.claude/claude-context-bar.py"
},
"subagentStatusLine": {
  "type": "command",
  "command": "python3 /YOUR_HOME/.claude/claude-agent-rows.py"
}

Configuration

Set these in the environment, for example inline in the statusLine command:

Variable	Default	Meaning
CCBAR_TTL_SECONDS	300	Prompt cache TTL in seconds. Use 3600 for a 1-hour cache; this also switches the write price to 2×.
CCBAR_BAR_WIDTH	24	Progress bar width in cells.

"statusLine": {
  "type": "command",
  "command": "CCBAR_TTL_SECONDS=3600 python3 /YOUR_HOME/.claude/claude-context-bar.py"
}

An invalid value falls back to the default.

How it times the cache

The cache is timed from the transcript file's last modification, which tracks the last request. Anything else that writes to the transcript also looks like a renewal. If the transcript can't be read, the ⏱ segment is left out and the rest of the bar still draws.

Subagent rows

The main bar always describes the main session, even while you view a subagent. claude-agent-rows.py, set as subagentStatusLine, adds each running subagent's model to its row in the agent panel:

○ general-purpose  Committing porSpan clamp fix             Opus 5.5 · 2m 7s · ↓ 110.9k tokens

Finished agents, and any row it can't read, keep Claude Code's default rendering. It works alongside either form of the bar.

Prices

Model	Input $/MTok	Cache read	Cache write, 5m (1.25×)	Cache write, 1h (2×)
Fable 5.1 / Mythos 5.1	10.00	0.25 (0.025×)	12.50	20.00
Fable 5 / Mythos 5	10.00	1.00 (0.1×)	12.50	20.00
Opus 5.5	4.00	0.20 (0.05×)	5.00	8.00
Opus 5 / 4.8 / 4.7 / 4.6 / 4.5	5.00	0.50 (0.1×)	6.25	10.00
Sonnet 5	2.00	0.20 (0.1×)	2.50	4.00
Sonnet 4.6 / 4.5	3.00	0.30 (0.1×)	3.75	6.00
Haiku 4.5	1.00	0.10 (0.1×)	1.25	2.00

From the Anthropic pricing page. An unrecognized model gets the timer without prices.

Two things make the prices approximate:

• The whole context is priced as cacheable. Only the part up to the last cache breakpoint is actually written, so the write price runs slightly high.
• Fast mode isn't priced. It shows the standard rate.

Keeping prices current

python3 check_pricing.py

This compares the table in claude-context-bar.py against the live pricing page and lists every figure that differs, including current models with no entry. It exits 0 when in sync, 1 on drift, and 2 if the page can't be fetched or read. A weekly GitHub Actions job runs it and opens a pricing-drift issue when prices change, closing it once they match again.

The mod keeps its own copy of the table in mod/hooks/format.ts. Update both together; test_mod_prices.py fails when they differ.

Development

python3 test_context_bar.py
python3 test_check_pricing.py
python3 test_agent_rows.py
python3 test_mod_prices.py
claude plugin test mod
claude plugin validate mod

The Python tests use only the standard library. The mod's tests run in Claude Code's plugin test kit on a mocked clock. Once the mod has loaded, Claude Code writes its API types to mod/.claude-plugin/types/ (gitignored), and npx -p typescript tsc -p mod type-checks it.