m3l-output-governor

m3l-output-governor is a Claude Code plugin that measures where Claude Code's output tokens go, so the levers that lower them can be chosen from your own sessions instead of guessed. It never changes what Claude sees, says or does.

Who is this for

• Someone who wants to lower Claude Code's output-token cost and would rather see their own numbers than follow a rule of thumb.
• Someone deciding whether to change the effort level and wants to know how much of their output is thinking first.
• Someone weighing the built-in Concise output style and wants the measured effect on a real task set, with its limits.

Quickstart

You need Claude Code (see Requirements for versions).

1. Install the plugin from the monte3l catalog. From your shell:

   claude plugin marketplace add monte3l/m3l-plugins
   claude plugin install m3l-output-governor@monte3l
   

   Or inside a Claude Code session:

   /plugin marketplace add monte3l/m3l-plugins
   /plugin install m3l-output-governor@monte3l
   

   To add the marketplace and install in one step, use claude plugin install m3l-output-governor --marketplace monte3l/m3l-plugins (Claude Code 2.1.292 or newer) or /plugin install m3l-output-governor --marketplace monte3l/m3l-plugins (2.1.275 or newer).

   The shell form installs at user scope unless you pass --scope project or --scope local. The session form opens the plugin's details and asks you to pick a scope.

   If you installed it before 1.0.0 from the monte3l-output-governor marketplace, remove that marketplace first, so the plugin is not installed twice:

   claude plugin marketplace remove monte3l-output-governor
   

   /plugin marketplace remove monte3l-output-governor
   

   The plugin name does not change. Its id becomes m3l-output-governor@monte3l.

2. In a project you have used with Claude Code, run /m3l-output-governor:measure. Claude reports the numbers for that project's own session transcripts.

3. Read thinkingShareOfOutput against Levers to see which one applies to you. On the author's sessions it was 36%.

Settings

This plugin declares no options, so there is nothing to configure at install. The settings that change your output tokens belong to Claude Code itself, and Levers lists them.

Manage the plugin

Task	From your shell	Inside a session
Update	claude plugin update m3l-output-governor@monte3l	/plugin, then Installed, the plugin, Update now
Update automatically	Not available	/plugin, then Marketplaces, monte3l, Enable auto-update
Turn off or on	claude plugin disable m3l-output-governor@monte3l, then claude plugin enable m3l-output-governor@monte3l	/plugin disable m3l-output-governor@monte3l, then /plugin enable m3l-output-governor@monte3l
Uninstall	claude plugin uninstall m3l-output-governor@monte3l	/plugin uninstall m3l-output-governor@monte3l
Run a checkout for one session	claude --plugin-dir /path/to/m3l-output-governor	Not available
Run an older release	Check out its tag, m3l-output-governor--vX.Y.Z, then use --plugin-dir	Not available

monte3l is not an official Anthropic marketplace, so Claude Code does not update its plugins automatically unless you turn auto-update on. Claude Code keeps you on the installed version until the plugin's version changes. See versions and updates. After a change inside a session, run /reload-plugins to apply it.

How it works

The plugin ships one skill and no hooks, so it changes nothing about how Claude Code runs. The skill runs scripts/measure.mjs --project on the current project's session transcripts, and Claude reports the numbers with the lever they support. The script only reads transcripts and prints counts: never prompt text, tool output or paths. The skill is user-invoked only, because it reads your session history.

That one call is everything the plugin runs. It has no bin/ directory, no MCP server and no hook, and the skill's allowed-tools pre-approves only node "${CLAUDE_PLUGIN_ROOT}/scripts/measure.mjs" *. Claude Code's own guidance is that an installed plugin can run code with your privileges, so read the script before you install. It has no dependencies and makes no network call.

Every number is labelled exact (read straight from a transcript's usage fields) or estimated (derived from characters, at 4 characters per token). See Measure from the command line for the full field list.

[!NOTE] An earlier design also shrank tool results with a hook. A baseline over 359 real sessions showed Bash results of 8,000 characters or more carry about 2% of billed input, so that layer was dropped. That baseline read each session's main transcript only; counting subagent transcripts too gives 2.4%. The numbers are in UNVERIFIED.md.

This is the beginning of the README. Read the rest at the original. Read the full README ↗