ccbs: open-source terminal dashboard tracks your Claude Code 5-hour and weekly limits live
jeayese · reddit · 2026-11-01
Reddit user jeayese released ccbs, a free open-source (GPL-3.0) btop-style terminal dashboard that shows Claude Code usage in real time so you never hit the 5-hour or weekly limits unexpectedly.
Features:
- 5-hour window: token graph, % of limit used, reset countdown, and a pace line predicting whether you'll hit the cap before reset
- Weekly: weekly limit %, reset time, daily tokens for the past week
- Now: each running session's project, branch, model, latest prompt, and last tool call
- History: duration, tokens, and API-equivalent cost for every past session
Data is parsed locally from Claude Code's transcript files in /.claude/projects; the 5h/weekly percentages come from the same endpoint as Claude Code's /usage, talking only to api.anthropic.com, with an offline estimation mode available.
The whole tool was built with Claude Code (Opus), mostly from the author's phone via a cloud session. Notable moments: the model self-checked its TUI via headless Chromium screenshots, found a startup hang through pty testing and replaced the framework, reproduced a Windows slow-load bug with 1.2 GB of fake history then optimized first frame to 20 ms, and fixed contrast across all 42 btop themes by computing ratios.
Key lesson from the author: giving the model ways to verify its own work (screenshots, benchmarks, pty tests) beats describing bugs from memory.
More from coding & agent
- Indie game made with heavy Claude help launches free demo on Steam — sidahuj · 2026-11-01
- Dev builds free Meta Muse alternative: an AI agent with its own browser that acts on your behalf — moorish-prince · 2026-11-01
- Proactive agents explained: the trick is putting approval gates only on irreversible actions — Arindam_1729 · 2026-11-01
- terrio v0.6.0: pure-Rust geospatial I/O without GDAL install, usable in browser and by AI agents — giswqs · 2026-11-01
- Mistral Large 4's Terminal-Bench 41% only holds inside Mistral's own agent — julsimon · 2026-11-01
- Anthropic research lead: prompting is over, 99% of engineers run 300+ self-improving agents — pstAsiatech · 2026-11-01