local-first · MIT · no account · no telemetry

Local usage analytics for
AI coding agents.

When an AI coding limit hits mid-session, work can stop. If you use more than one agent, usage is split across separate logs and provider views. Tokimeter turns the records already on your machine into one private report of tokens, API-equivalent cost, models, projects, and honest limit signals—before a cutoff surprises you. Supports Claude Code and Codex across CLI and desktop sessions, plus Cursor, Grok Build, Hermes, opencode, Cline, and Copilot CLI. No account. No CLI telemetry. Prompts never leave your machine.

GitHub

npx tokimeter report

$ npx tokimeter report

claude~$0.00

codex~$0.00

grok~$0.00

hermes~$0.00

total ~$0.00 at API rates ✓ read locally · 0 bytes sent

working CLI flow · run npx tokimeter report now · synthetic numbers shown, your report reads your real local records

Quick reportone-time · no setup

$ npx tokimeter report

Install Tokimeterongoing · stays local

$ npx tokimeter install

Star on GitHub

Requires Node.js 18+, which includes npm and npx. Run a one-time report first, or install Tokimeter for an ongoing local meter.

// working product

Run it on your own sessions in one command.

Tokimeter is published on npm, open source on GitHub, and usable without creating an account. The zero-setup report reads supported agents' existing local records immediately.

View npm package Inspect the source

01 · Runnpx tokimeter report
One-time report with no installation, account, wrapper command, or provider credential required.
02 · Read locallyExisting agent records
Tokimeter parses numeric usage metadata already written by supported coding agents. It does not read prompts, responses, source code, or diffs.
03 · See the answerOne cross-tool report
Break down usage by tool, provider, model, project, and day, with reported, estimated, and unknown pricing kept visibly separate.

How Tokimeter differs from the usual options
Capability	Provider dashboards	Tool-specific usage readers	Tokimeter
Coverage	That provider's products	Usually one coding tool	Eight coding-agent tools in one report
Cross-tool view	No	Usually no	Yes—tool, provider, model, project, and day
Local-only mode	No	Varies	Yes—the free CLI needs no account or telemetry
Cost honesty	Provider-specific	Varies	Reported, API-equivalent estimated, and unknown values stay separate

// limits

Am I about to hit my limit?

npx tokimeter limits groups locally observed usage into 5-hour and 7-day views for budgets you set. When a tool records vendor counters, Tokimeter shows those separately; it never invents a provider quota or reset time.

5h window

resets in 1h 9m

vendor-reported

Weekly cap

resets Mon 05:46

⚠ pace yourself

tokimeter limits

  claude
    Last 5h:   ~$49.91 · 149 calls · 40.2M tokens · 83% of ~$60 budget  ⚠
    Last 7d:   ~$49.91 · 149 calls · 40.2M tokens

  codex
    Last 5h:   ~$11.96 · 109 calls · 22.6M tokens
    Vendor-reported (plus plan), as of 11:18:
      5h window:  48% used · resets in 1h 9m (13:08)
      Weekly:     73% used · resets in 4d 17h (Mon, Jul 13 05:46)

01 Real reset times, not guesses

Codex records OpenAI's own rate-limit counters locally; Tokimeter surfaces them with exact reset times. Claude Code doesn't log rate limits, so you get the honest rolling window instead of an invented countdown.

02 Budgets & warnings

Daily, weekly, per-tool 5h, and per-session budgets, with warnings in the CLI and in Claude Code's own status line once you cross 80%. Never blocking.

03 Keep your commands

Daily use stays claude and codex. No wrapper commands to remember.

// analysis

More than a total

The same local data answers the questions a running total can't. Every one of these is computed on your machine, with no extra model calls.

tokimeter savings

Where the bill could shrink

Finds routine work running on a premium model and shows what the same tokens would cost on a cheaper tier. It's an upper bound, stated plainly, never a nudge.

tokimeter agents

Director vs subagents

Splits a session into the main thread and the subagents it spawned, with cost broken out by agent type and the skills they used.

tokimeter plan

How much headroom is left

How much of each budget window remains, in dollars and tokens, and how long until you reach it at your current pace.

tokimeter trace <session>

One session, explained

A single session end to end: cost, the models it used, how work was delegated, and whether its cache actually paid off.

tokimeter report --orchestration

Which tools work together

The projects where you run two tools in the same window, and how the spend splits between them. Correlation, shown as correlation.

tokimeter report --md · --html

A report you can hand off

A clean Markdown or HTML report with per-project lines, ready for a client chargeback or your own records.

// honest numbers

Honest numbers, by design

Exact where the data is exact, and clearly marked ~ where it's an estimate.

Exact token counts

Read straight from the same local files your tools already write, including Claude Code and Codex desktop coding sessions. Nothing is sampled or modeled.

~$X means estimate

On a subscription you pay a flat fee. The ~$X is what that same usage would cost at API rates, so you can see the value you're pulling out. Provider/tool-reported costs keep their provenance; paid proxy accounting remains experimental until invoice reconciliation.

Four buckets, visible provenance

Input, cache write, cache read, and output are priced separately. Every price is labeled verified built-in, community feed, custom local, or provider/tool reported. Unknown models stay out of priced totals and appear only as a separate rough estimate.

No spend to save spend

Every insight and tip is computed locally from your own metadata. Tokimeter never makes a model call of its own.

// editor

In your editor

A VS Code / Cursor extension shows today's spend in the status bar and a cost dashboard inside the editor, using the same local metadata with no account. It's a viewer for the local tool; the CLI stays the source of truth.

Download tokimeter-vscode-0.1.0.vsix.
In VS Code or Cursor: Extensions panel → ⋯ menu → Install from VSIX… (or run code --install-extension tokimeter-vscode-0.1.0.vsix)
Run Tokimeter: Show Cost Dashboard from the command palette.

Download .vsix

Works in VS Code, Cursor, and other VS Code forks. Release notes and the verified checksum are available on GitHub.

Tokimeter Dashboard ×

Tokimeter

sample local data

AllClaudeCodex

Today

~$54.32

221 calls

Recent total

~$367.37

2,953 calls

5h window

48%

resets 1h 9m

Latest

gpt-5.5

high · exact

Today by tool

codex~$38.42109 calls

claude~$11.9687 calls

grok~$3.1018 calls

hermes~$0.847 calls

⑂ main ✓ local only◉ Tokimeter today ~$54.32 · 5h 48%

// security

Security & privacy

Tokimeter stores usage metadata only: tokens, models, costs, project paths. Never prompt or response content, because there is no field for it anywhere, by construction. Nothing is sent anywhere by default, so no telemetry and no phone-home. The optional proxy binds to localhost and never logs your API keys.

Read the full security notes →

// pro

Tokimeter Pro

The local tool is free and complete. Pro and Team add what a local tool can't: history that outlives your logs, alerts that reach you away from the terminal, and a shared view for a whole org.

Local

The local CLI and reports, always free.

✓ Full local reports and demo
✓ Data stays on your machine
✓ No account or telemetry

Install locally

Install Tokimeter in one command.