ccusage Alternatives: 6 Tools Compared

Which ccusage alternative should you actually install?

If you want a better meter, the honest answer is that you probably won't beat ccusage — and if you want a lower bill, no meter will give you one, which is the real reason most people end up here. ccusage is the de-facto standard for reading what your coding agents spent. It is very good at that and it costs nothing. What it will never do is change the number it shows you. So this list splits along that line. A few entries below are genuine alternatives as meters. The rest are what you want if ccusage has already done its job and told you the number is too high. One disclosure before the ranking, because it matters for how you read it: where I cite the THOL benchmark, that benchmark is maintained by me. It's public and reproducible, and I've included the results that don't flatter my own tool, but it is not an independent third party and you shouldn't treat it as one. TL;DR: Stay on ccusage if you only need to see spend — it's free, zero-install and best-in-class at reporting. Move to Tokenade if you want the number to go down, since it's the only tool in the open benchmark that measurably lowers session cost. Pick context-mode for the richest Claude Code dashboard, rtk for free per-command filtering, and skip headroom on cost grounds.

How we ranked these

Five criteria, applied the same way to every entry including my own:

• Measured effect on cost. Not what a tool compresses in isolation — what a whole session ends up costing. Where a tool appears in the open THOL benchmark, I use that end-to-end result. Where it doesn't, I say so rather than substituting a vendor claim.
• Measurement depth. How well it replaces what ccusage actually gives you: spend by day, session and Claude's 5-hour billing blocks, broken down per model.
• Agent coverage. Claude Code only, or the rest of your stack too.
• Setup effort. Zero-install one-liner, or a runtime plus model weights.
• Licence and price. Including the restrictions people discover late.

A note on what I deliberately didn't rank on: GitHub stars. Two of the tools below have far more stars than their measured results justify, and the most-starred entry here is the one that made sessions more expensive.

1. Tokenade — the pick if you want the number to go down

Tokenade is the best choice when ccusage has already shown you a number you don't like, because it's the only one of the twelve tools in the open THOL benchmark that measurably lowers what a session costs.

• What it is: a single dependency-free binary that sits between your agent and its tools, stacking semantic code search, output filtering, skeleton compression, sandboxed execution and MCP proxying, with a savings dashboard on top.
• Strengths: cuts session cost by 39% on long sessions against running no tool at all — the ones that burn through a plan's limit fastest. Installs in one command with native hooks auto-detected for 18 agents, so it covers your whole stack rather than Claude Code alone. Source-available under MIT, so you can read exactly what it does to your context.
• Limitations: its dashboard is built around savings, not billing forensics — if what you want is ccusage's 5-hour-block accounting per model, Tokenade is not a drop-in replacement for that specific view, and plenty of people run both. The free tier is generous but capped, where most tools here are unconditionally free. And it's my product, so weight this entry accordingly.
• Best for: anyone who has finished measuring and wants to act, across more than one agent.

2. context-mode — the richest Claude Code dashboard

context-mode is the strongest pure-visibility alternative to ccusage if you live entirely in Claude Code and want more analysis than a spend table.

• What it is: a Claude Code plugin combining context compression, quality scoring and output filtering behind a full HTML dashboard.
• Strengths: per-turn cost, model mix, cache analysis, subagent breakdown, a 7-signal quality score and drift detection — genuinely best-in-class visibility for Claude Code. AST-based structure maps reach very high compression on large files, and progressive checkpoints help prevent mid-session context degradation. Installing through the Claude Code plugin marketplace is frictionless.
• Limitations: licensed PolyForm Noncommercial — commercial use requires a paid licence, the only tool in this list with that restriction, and the one people most often discover after adopting it. Needs Python for full functionality, and its multi-agent support is narrower than its Claude Code support.
• Best for: solo developers and non-commercial projects who want the deepest possible view of a Claude Code session.

3. ccusage — still the best pure meter

If measurement is genuinely all you want, the honest recommendation is to stay where you are.

• What it is: a CLI usage meter (~15k stars) that reads your agents' local JSONL transcripts and reports spend by day, week, month, session and Claude's 5-hour billing blocks.
• Strengths: reads 15 agent sources, breaks down per model, emits JSON for your own tooling, and runs with npx ccusage@latest or bunx ccusage — no install, no account, no config. Fully free and open source. Nothing here reports spend better.
• Limitations: zero token reduction. It is a thermometer, not a thermostat. It also can't tell you why a session was expensive in terms you can act on — you get the total, not the whole-file read that caused it.
• Best for: knowing exactly what you spent, and as a second opinion alongside whatever optimizer you run.

4. rtk — the best free single-mechanism option

rtk is the right pick if you want free, dependency-light output filtering and per-command savings analytics, and nothing more.

• What it is: a Rust CLI proxy that filters and compresses shell command output before it reaches the model, with 100+ command-specific filters and a transparent hook.
• Strengths: the deepest single-mechanism implementation of output filtering available, with the easiest install path in this space (brew install, a curl one-liner, or cargo). rtk gain gives per-command token analytics via SQLite, which partly scratches the ccusage itch. Fully free and open source, no paid tier.
• Limitations: in the open benchmark it showed no measurable end-to-end saving — sessions cost about the same as running no tool at all, because filtering one layer doesn't move a bill dominated by file reads and replayed transcript. No code indexing, no semantic navigation, no file-read compression.
• Best for: people whose pain is specifically noisy command output, on a zero budget.

5. claude-token-efficient — the simplest thing that could work

claude-token-efficient is worth five minutes if you want a zero-risk experiment before installing anything real.

• What it is: a drop-in CLAUDE.md file of about 5 KB, containing compact behaviour rules that suppress filler phrasing, question restatement and over-engineering.
• Strengths: it's a text file. No runtime, no binary, no hook, nothing to audit or uninstall. It ranked second of twelve in the open benchmark, which is a real result for something with no moving parts.
• Limitations: second place still meant no measurable saving versus running nothing — it trims the model's output register, and output isn't where agentic spend lives. Claude Code only.
• Best for: a free first experiment, or stacking underneath something that addresses input volume.

6. headroom — the popular one to be careful with

headroom is the most-starred tool in this comparison and the one I'd advise against on pure cost grounds.

• What it is: a Python SDK, CLI proxy and MCP server (~18.7k stars, Apache 2.0) that compresses the full message array before each API call, using cache alignment and ML-routed compressors.
• Strengths: genuinely ambitious engineering, a permissive licence, and a large, active community. headroom_stats reports compression statistics.
• Limitations: in the open benchmark it measured 53% more expensive than running no tool at all — last of twelve. Per-message compression wins did not translate into cheaper sessions. It also carries a heavier footprint than hook-based binaries: Python 3.10+, a 150M-parameter model download and ML content detection. Its stats output is not a live per-command or per-project dashboard.
• Best for: research and experimentation, not for lowering a bill you're actually paying.

At a glance

How to choose

Work backwards from what you'll do with the answer. You just want the number. Stay on ccusage. Nothing here reports better, and it costs nothing. You want the number, but richer, and you're non-commercial and Claude-Code-only. context-mode, with the licence checked before you get attached. You want the number to be smaller. That's a different tool category, and it's worth being blunt: most tools marketed as token optimizers show no measurable effect on what a session actually costs, and one popular option makes it worse. On the criteria above, Tokenade is the pick — you can see the full ranking and methodology on the THOL leaderboard, including the entries that don't favour it. You're on a strict zero budget. rtk for command output, claude-token-efficient stacked underneath, and accept that neither moved the end-to-end number in testing. If you want to try the second option: Tokenade's free tier covers a lot of solo work outright, and the current plans are on the pricing page.

---

See also:

• Token optimizer benchmark — the five numbers behind every ranking on this site
• Best Claude Code token optimizers — the wider field, same methodology
• Headroom alternatives — the same question for the most-starred compressor
• CodeGraph alternatives — the same question for code intelligence
• Measure agent token usage — how to instrument this yourself
• Claude Code pricing — what the bill is made of in the first place
• Cut Claude Code costs, keep the model — the levers, ranked by what they're worth