See your AI coding costs in real time. Tokenline adds context usage, prompt-cache savings, TTL countdown and rate-limit pacing to Claude Code and Gemini CLI.
If you have Node 18+, run this inside your project repository to configure the statusline locally.
npx @inbrace-tech/tokenline initWant it globally for all your projects? Add --global to the command.
npx @inbrace-tech/tokenline init --globalSimply restart Claude Code or Antigravity, and your new statusline will be live.
- Line 1 — model · context used (tokens + %) · cache TTL with a live HOT→COLD countdown.
- Line 2 — per-turn token economics:
read / write / new / output, the equivalent billed tokens (eq), and thesaving %you get from prompt caching. - Line 3 — 5h and 7d rate-limit bars with reset ETA and a pace marker (
!!= you're burning the window faster than it refills).
Lines 2 and 3 appear only when there's something to show (a turn happened, limits exist), so the bar stays quiet when idle.
tokenline turns the one-line status bar of your AI coding CLI into a live cockpit:
- Which model you're on.
- How much context you've burned.
- How long your prompt cache stays hot.
- How many tokens you're saving by reusing it.
- How close you are to your 5h / 7d rate limits.
It is cross-CLI (Claude Code, Antigravity) and cross-provider (Anthropic, Gemini) — detecting the active client and model provider at runtime and adjusting the cost equivalents accordingly.
Most statuslines show the model and the context bar. tokenline adds the two things that actually drive cost and flow on long agent sessions:
- Cache visibility. Anthropic and Gemini bill cached input tokens at a fraction of the price — but the cache expires.
tokenlineshows the TTL countdown (5m or 1h, detected from the data) so you know whether your next turn lands warm or cold. - Savings, quantified. The
saving %makes the value of prompt caching concrete instead of invisible. - Rate-limit pacing. The
!!marker warns you when you're on track to hit the 5h or 7d ceiling before it resets.
LLMs are stateless — by default, they must reread your entire codebase context (system prompt, tools, file history) on every single turn before generating a response. This phase (Prefill) is slow and expensive.
Prompt Caching solves this by saving that processed state in the provider's memory:
- Cache Write (Cold): The model reads everything and stores the state. This costs slightly more than base tokens.
- Cache Hit (Warm): If your prompt prefix matches the cached state exactly, the model skips the reading phase. This is ~90% cheaper and starts responding almost instantly.
- TTL (Time-To-Live): The cache operates as a sliding window. Every cache hit resets the countdown to its full duration for free. The secret to minimizing rate-limit drain is adjusting your workflow pace to keep the cache continuously HOT.
Anthropic's rules for assigning cache TTLs are dynamic and can change without notice (e.g., historically granting a 1-hour TTL to Claude.ai Max users). This is exactly why tokenline is valuable: it doesn't guess your TTL based on your plan. It parses the actual CLI telemetry in real-time to detect whether your last turn was processed as a 5-minute or 1-hour write, and adapts the UI countdown and token economics accordingly.
If tokenline reveals you are getting 1-hour cache writes, keep in mind they burn 2x the tokens from your 5h/7d rate limit (compared to 1.25x for a 5-minute write). If you find this is exhausting your quota too quickly—especially when running multiple subagents that cool down fast—you can force the CLI to fall back to the cheaper 5-minute writes by adding this to your .claude/settings.json:
"env": {
"FORCE_PROMPT_CACHING_5M": "1"
}v1 targets Linux / WSL2:
bash4 or newerjq- GNU coreutils (
date -d,stat -c)
macOS and Windows support are on the roadmap.
install.shchecks all of the above and tells you exactly what's missing.
No Node? Clone the repo and run the dependency checker, which prints a ready-to-paste snippet:
git clone https://github.com/inbrace-tech/tokenline.git
cd tokenline
./install.shAdd the printed block to your project's .claude/settings.json (or ~/.claude/settings.json for global), inside the top-level object:
"statusLine": {
"type": "command",
"command": "bash /absolute/path/to/tokenline/tokenline.sh",
"refreshInterval": 1
}Then restart Claude Code.
npx @inbrace-tech/tokenline init is deliberately transparent about touching your config:
- Writes
tokenline.shto./.claude/(or~/.claude/with--global). - Merges only the
statusLinekey intosettings.json— every other setting is preserved. - Backs up
settings.jsontosettings.json.bakbefore writing. - Never clobbers invalid JSON: if it can't parse your
settings.json, it stops and prints the block to paste manually. - Is idempotent, and won't replace a different existing
statusLineunless you pass--force.
Other commands: doctor (check dependencies and config, change nothing) and uninstall (remove the block; --purge also deletes the script).
tokenline detects the Antigravity CLI from the transcript path and switches to its provider equivalents automatically. Point Antigravity's statusline command at the same
tokenline.sh — no extra flags needed.
On every refresh the host CLI pipes a JSON snapshot of the session to the script over stdin. tokenline parses it in a single jq pass (to keep the per-second refresh cheap), reads the last turn's timestamp from the transcript to drive the cache countdown, and renders up to three lines. Per-turn timestamps are cached in a per-user 0700 directory under $XDG_RUNTIME_DIR (tmpfs, cleared on logout).
| Symptom | Cause / fix |
|---|---|
Blank statusline, or [tokenline] jq not found |
Install jq (apt install jq / brew install jq), then re-run ./install.sh. |
Cache shows COLD immediately |
Normal right after a long idle gap — the cache window has elapsed. It goes HOT again on your next turn. |
| Colors look wrong / show escape codes | Your terminal must support 256-color ANSI. Most modern terminals do; check your $TERM. |
| Nothing renders on macOS | Expected on v1 — macOS uses BSD date/stat. See the roadmap. |
- macOS support (BSD
date/stat, bash 3.2 fallback) - Windows support (Git Bash / PowerShell)
- Configurable colors and thresholds via
TOKENLINE_*env vars
(Issues for these are tracked in the repo — contributions welcome.)
Issues and PRs are welcome — see CONTRIBUTING.md. The CI runs ShellCheck on every push, so please keep the script lint-clean.
tokenline is built and maintained by Inbrace — a software house based in Campinas, Brazil, building software with technical and human responsibility. We work where security is non-negotiable and architecture is treated as a strategic asset: intentional engineering, grounded decisions, and systems built to last.
Inbrace. The human side of software.
Learn more at inbrace.com.br · LinkedIn
MIT © 2026 Inbrace.