Local-first / Open source

Optimise your coding agents.

See what every agent run costs while it runs: tokens, estimated spend, context, cache, and tool calls for Claude, Codex, Cursor, OpenCode, Kiro, Pi, and Antigravity. Read from the traces already on your machine.

  • No API keys
  • No telemetry
  • Python standard library
  • Serves on 127.0.0.1
Start Install locally
$
git clone https://github.com/splunk/token-meter.git
./token-meter/scripts/install

Requires Python 3.8+ · Xcode Command Line Tools · curl

Then open http://127.0.0.1:8722 No administrator access required.

Things your bill won't tell you.

Surfaced by Token Meter on one engineer's laptop, estimated at public list prices.

Read the five levers
  1. 6% → 55%

    Cache writes were 6% of tokens and 55% of the dollars on Claude Opus 5.5.

  2. 83,907

    Tokens of tool definitions loaded into prompts and never called.

  3. $6.04 → $0.96

    Cost per subagent run, a month after switching to short, narrowly scoped tasks.

What you'll see.

One local dashboard at 127.0.0.1:8722. Costs and selected token values can be estimates, and coverage stays visible. Screens show synthetic sample sessions.

Token Meter Sessions view with four live session cards showing estimated cost, output per dollar, context in use, and speed
Live sessionsCost, context, cache, and speed for every running agent.
Menu-bar companion A fast read without opening the dashboard.

Rebuilt here in HTML. The real one sits in the macOS menu bar, the Linux tray, or the Windows notification area.

What becomes visible.

Live runs

Cost, context pressure, cache, pace, and the latest execution for every running session.

SESSIONS / MENU BAR

Spend

Estimated spend by day, runtime, project, and session, with machine-wide monthly budgets.

SPEND / BUDGETS

Models

Cost, speed, cache, and context per model, scoped by runtime so identical names stay separate.

MODELS / EFFICIENCY

Delivery

Pushed lines from local Git beside covered spend, plus tool, MCP, and skill evidence without contents.

GIT / TOOLS

Efficiency, with the evidence beside it.

How Token Meter reads efficiency

  • Output / covered $

    Output tokens per covered dollar. Higher is more output for the spend.

  • Reasoning ratio

    Reasoning share of output. Stable is usually better; hard work may need more.

  • Context load

    Input tokens per output token. Lower carries less context into each response.

  • Cache hit ratio

    Share of cache-covered input read from the prompt cache. Higher is usually better.

No universal score. Read movement across comparable work. Coverage remains visible: missing evidence never becomes zero. Costs and selected token values can be estimates.

Open source from Signals, not verdicts.

Your traces stay on your machine.

Token Meter reads supported local stores and serves the dashboard on 127.0.0.1. It is not a hosted analytics service.

Inspect the security boundary
Read-only inputs

Supported local traces and databases are read, never rewritten.

Bounded output

No prompt or response content appears in the dashboard, native companion, or bounded MCP projections.

No hosted plane

No API keys for trace analysis. No Token Meter analytics or telemetry leaves your machine.

Your agents are already producing the evidence.

Make it useful. Keep it local.

Install Token Meter