MYCELLIUM
Menu
Start a project

Open source · Token reducer · Public beta

TokenPilot

Cut the tokens your AI coding CLI spends. Same model. Same tools.

  • 76.50%fewer tokens, best resultGrok 4.7 Build Fast
  • 4 of 4Grok Build modelsat or above 42% fewer tokens
  • 6 of 9OpenAI Codex modelsat or above 42% fewer tokens

Oct 6, 2026 · one paired run per model on the same task, every check passing · total input plus output, cached input included · not a universal estimate

npm install -g @cesaremcasa/tokenpilot@0.5.6GitHub

Benchmark

Measured on a real task, checks still passing.

Token reduction per model, October 6, 2026. Grok Build: grok-4.7-build-fast 76.50%, grok-4.7 65.49%, grok-4.6 57.89%, grok-4.5 57.10%. OpenAI Codex: gpt-daybreak-blue-latest 62.24%, gpt-6-sol 61.62%, gpt-5.5 54.33%, gpt-5.6-luna 54.29%, gpt-5.6-sol 48.17%, gpt-6-luna 47.63%, gpt-6.1-sol 26.34%, gpt-5.6-terra 24.14%, gpt-6-astra 19.78%. One paired run per model; not a universal estimate.Token reduction per model, October 6, 2026. Grok Build: grok-4.7-build-fast 76.50%, grok-4.7 65.49%, grok-4.6 57.89%, grok-4.5 57.10%. OpenAI Codex: gpt-daybreak-blue-latest 62.24%, gpt-6-sol 61.62%, gpt-5.5 54.33%, gpt-5.6-luna 54.29%, gpt-5.6-sol 48.17%, gpt-6-luna 47.63%, gpt-6.1-sol 26.34%, gpt-5.6-terra 24.14%, gpt-6-astra 19.78%. One paired run per model; not a universal estimate.
OpenAI Codex · token totals per run · 6 of 9 models at or above 42%
ModelWithoutWithCut
gpt-daybreak-blue-latest273,252103,18262.24%
gpt-6-sol329,492126,45961.62%
gpt-5.5279,125127,47454.33%
gpt-5.6-luna228,427104,41754.29%
gpt-5.6-sol200,647103,98748.17%
gpt-6-luna338,034177,04047.63%
gpt-6.1-sol139,360102,65126.34%
gpt-5.6-terra211,652160,55524.14%
gpt-6-astra170,126136,47119.78%
Grok Build · token totals per run · 4 of 4 models at or above 42%
ModelWithoutWithCut
grok-4.7-build-fast831,211195,32376.50%
grok-4.7643,271222,01665.49%
grok-4.6298,629125,76557.89%
grok-4.5512,361219,81257.10%

One pair per model, not a statistical guarantee; everyday workloads may differ. Claude, Kimi and Grok Bot are unproven. Evidence files.

How it runs

A local launcher in front of the CLI you already use.

How TokenPilot runs: you run your CLI as usual; the local TokenPilot launcher checks the CLI version; if every option is supported it applies a versioned session policy; your CLI (Claude Code, Codex or Grok Build) runs with the same model and tools; the provider's published counters feed a local report. If support cannot be confirmed, nothing is injected and the CLI starts unchanged.How TokenPilot runs: you run your CLI as usual; the local TokenPilot launcher checks the CLI version; if every option is supported it applies a versioned session policy; your CLI (Claude Code, Codex or Grok Build) runs with the same model and tools; the provider's published counters feed a local report. If support cannot be confirmed, nothing is injected and the CLI starts unchanged.

Usage

Install once. Keep typing the same commands.

Supports Claude Code, Codex and Grok Build on macOS and Linux. Kimi runs through unchanged.

# install the launchers, then check setup
$ tokenpilot install
$ tokenpilot doctor

# use your CLI exactly as before
$ codex --model gpt-5.5

# see this session's numbers
$ tokenpilot report --provider codex

# skip TokenPilot any time
$ TOKENPILOT_BYPASS=1 codex

Why we built it

Stop paying for tokens that do nothing.

Agents reread files, replay context and think longer than the task needs, and every one of those tokens is billed. TokenPilot is how Mycellium closes that leak in its own work. Now it is open source, so you can close it in yours.

Private by design

No proxy. No prompts, responses, code or credentials stored. Local, content-free counters only. tokenpilot mode off turns it off.

Limits

Public beta. macOS and Linux only. Session cache figures are not savings, and token usage is not a provider bill. MIT licensed; not affiliated with Anthropic, OpenAI, xAI or Moonshot AI.

Open source

Read the code. Try it on your own sessions.