Seven tools that tell you what Claude Code is costing you — and the one difference that matters most: some only report the spend, and some can actually stop it. Written by the makers of Terse, which is one of the seven. We say plainly where the others are the better pick.
| Tool | Form | Price | Pick it when |
|---|---|---|---|
| ccusage | npm CLI | Free, open source | You want the number and nothing else. |
| Claude-Code-Usage-Monitor | Python TUI | Free, open source | You live in the terminal and want a live readout. |
| TokenTracker | macOS / Windows app | Free | You run many different AI tools and want one ledger. |
| CCSeva | macOS menu bar | Free, open source | You want a glanceable number in the menu bar. |
| Anthropic Console | Web | Included | Reconciling what you were actually billed. |
| LiteLLM | Gateway / proxy | Free, open source | An org that needs one enforcement point across many people. |
| Terse | macOS / Windows app | Trial, then $4.99/mo | You want the bill changed, not just measured. |
Almost every tool in this category answers the same question — what did that cost? — and answers it after the money is gone. That is useful for budgeting and useless in the situation that produces most large bills: an agent looping overnight on a task it cannot finish.
Stopping spend is a different mechanism. It has to act before the next API call, which means enforcement at the process level or at a gateway. Only two tools here do that:
SIGSTOP-pausing or SIGTERM-killing the agent process before its next call.Everything else in this comparison is a reporting tool. That is not a criticism — a reporting tool that is free and installs in one command is the correct choice for most individuals.
Strength. The fastest honest answer to "what did this cost". Reads local JSONL, prints daily/weekly/monthly and per-session reports. No account, no proxy, no setup.
Limits. Reports only. No enforcement, no compression.
Pick it when: You want the number and nothing else.
Strength. Live terminal dashboard with progress bars, burn-rate analysis and predictions that warn before you hit a limit.
Limits. Warns, but cannot stop the agent. Claude Code only.
Pick it when: You live in the terminal and want a live readout.
Strength. Breadth: around 30 tools in one place, native apps and desktop widgets, correct de-duplication of token records across providers.
Limits. Reporting-focused; no spend enforcement or prompt compression.
Pick it when: You run many different AI tools and want one ledger.
Strength. A clean native menu bar readout of live Claude Code usage — always visible, never in the way.
Limits. Claude Code only, macOS only, reporting only.
Pick it when: You want a glanceable number in the menu bar.
Strength. The billing source of truth. Usage and cost by workspace, model and API key, with CSV export for finance.
Limits. Delayed rather than live, and it cannot see per-session or per-project detail.
Pick it when: Reconciling what you were actually billed.
Strength. Team-level control: every request routed through one proxy, with per-key hard budgets that reject calls once a limit is hit.
Limits. You must run and secure a proxy, and your prompts pass through it.
Pick it when: An org that needs one enforcement point across many people.
Strength. Acts on spend instead of only reporting it: a burn-rate/token/dollar ceiling that SIGSTOPs or SIGTERMs a runaway agent before its next API call, 40-70% on-device prompt compression, MCP risk scoring, and ~25 waste scans. Covers 8 agents.
Limits. Commercial after the trial. No Linux desktop build. Newer and smaller than ccusage.
Pick it when: You want the bill changed, not just measured.
Before choosing a tool, it helps to know what it will show you. Four things drive a Claude Code bill:
A tracker that reports only a daily dollar total will show you none of these. Look for cache read/write split, per-turn cost and context-window fill.
There is no single best one, because they do different jobs. For a free CLI that answers "what did Claude Code cost me this week", ccusage is the strongest choice — one command, no setup, reads the local JSONL. For a live terminal dashboard with burn-rate predictions, Claude-Code-Usage-Monitor. For coverage across many tools at once, TokenTracker. For a native macOS menu bar readout, CCSeva. For org-wide billing truth, the Anthropic Console. For gateway-level control across a team, LiteLLM. Terse is the one that acts on spend rather than only reporting it, with a budget circuit breaker that pauses or kills a runaway agent before its next API call.
Yes, several. ccusage is free and open source and runs with npx ccusage. Claude-Code-Usage-Monitor is a free open-source Python TUI. CCSeva is a free macOS menu bar app. The Anthropic Console is included with your account. Terse has a free 30-day trial and then costs $4.99/month, and its SDK and benchmark harness are MIT licensed and free.
Reporting tools cannot stop it — by the time a dashboard shows the number, the API call is already billed. Stopping spend needs enforcement at the process level: a ceiling on burn rate, tokens or dollars that pauses the agent process with SIGSTOP or terminates it with SIGTERM before the next API call fires. Terse is the tool in this comparison that does that. A gateway like LiteLLM can also enforce a hard budget by refusing requests once a key exceeds its limit, at the cost of running a proxy.
ccusage reports; Terse intervenes. ccusage is a free open-source CLI that reads Claude Code's local JSONL files and prints daily, weekly, monthly and per-session cost reports. Terse is a commercial desktop app that also monitors live, then adds a budget circuit breaker that stops a runaway agent before its next API call, compresses prompts 40-70% before they are sent, risk-scores MCP servers, and scans for duplicate tool calls and redundant file reads. If you only want the number, use ccusage.
The local-first ones do not send it anywhere. ccusage, Claude-Code-Usage-Monitor, CCSeva, TokenTracker and Terse all read the session files Claude Code already writes to ~/.claude/ on your own machine. Gateway tools are different by design: LiteLLM and similar proxies sit in the request path, so your prompts pass through infrastructure you have to run and secure yourself. The Anthropic Console shows billing metadata rather than prompt content.
Tools that read Claude Code's JSONL transcripts see the exact token counts the API reported, including the cache-read and cache-write split, so their token figures are accurate. Dollar figures are only as accurate as the price table each tool ships, so a stale table after a pricing change produces confident wrong numbers. The Anthropic Console is the billing source of truth; treat local tools as a fast, detailed estimate and reconcile against the Console.
Yes. TokenTracker covers roughly 30 tools and Terse covers eight agents — Claude Code, Cursor, OpenAI Codex, GitHub Copilot CLI, Cline, Windsurf, OpenClaw and Aider — in one view. ccusage has grown beyond Claude Code to several CLIs. Claude-Code-Usage-Monitor and CCSeva are Claude Code only. Gateways such as LiteLLM see every tool routed through them, regardless of vendor.
Terse's compression figures are reproducible rather than asserted. The benchmark harness is MIT licensed and in the repo:
git clone https://github.com/Terse-AI/terseai.git
cd terseai && npm run benchmarkIt reports text compression, working memory, tool optimization and model routing separately, because which lever helps most depends entirely on your workload. The full side-by-side, including when to use something else, lives in COMPARISON.md.
Terse monitors 8 agents live, compresses every prompt 40–70% on-device, and stops a runaway agent before its next API call. macOS & Windows. Free 30-day trial.
Related: Terse for Claude Code · Claude Code pricing 2026 · How to reduce AI API costs · What is a budget circuit breaker · MCP manager · AI coding agent costs