gauge
Measure whether a change helped a coding agent
TLDR
SYNOPSIS
gauge [command] [options]
DESCRIPTION
gauge is the command from the npm package `@withgauge/cli` (gauge-sh/gauge-cli). It runs coding agents against tasks you define, stores the transcripts, and gives you a way to see whether a later change did better. It is meant to be driven by the agent that is editing the project, as well as from the shell.Local mode needs Docker running Linux containers, plus credentials for Claude Code, Codex, Pi, or OpenCode. init --local keeps measurements in `.gauge/`, which should stay uncommitted. --cases adds Markdown cases you can commit. Skills and files are read from the working tree. An npm CLI or SDK under test should be packed first. Gauge installs that tarball in the sandbox.Hosted mode adds docs, authenticated MCP servers, and authenticated APIs. gauge auth login (or --device on a remote terminal) is required for hosted Chat and Agents. gauge onboard picks the workspace.`gauge instructions` is how an agent learns the current command surface. `gauge --help` and `gauge run --help` cover flags this page does not list.
PARAMETERS
local runtime pull
Download the Docker image pinned to this CLI release. It contains the agent harnesses Gauge runslocal runtime inspect
Show the local runtime Gauge will useinit --local
Create `.gauge/`, including a SQLite database. This does not change tracked filesinit --local --cases
Also write `gauge.json`, a `gauge-cases/` directory, and a note in `AGENTS.md`init --hosted
Point this checkout at the hosted backendinstall-skill
Install the Gauge skill so a coding agent can find the CLI's instructionsauth login [--device]
Sign in for hosted Chat and Agents. --device is the flow without a local browseronboard
Sign in and create or select a workspaceeval create --local --name name
Save an eval in the local databaserun --eval id [--local | --hosted] [--input name] [--dry-run] [--verify] [--files glob]
Execute an eval. --dry-run does not spend a full session. --files limits which case files are includedrun compare --local --baseline batch=id --candidate batch=id
Compare two local batchesrun analysis save
Record findings from a local session. Hosted mode analyzes sessions itselfrun insight
Show the evidence for a runrun list --local -o json
List local runs as JSONpreference create --kind open-ended
Start a choice study instead of a fixed evalinstructions [topic] [--local | --hosted] [--list]
Print the instructions a coding agent should follow. Topics include init, run, eval, preference, and session-analysis
CONFIGURATION
.gauge/
Private workspace: SQLite database, backend preference, and local measurements. Do not commit itgauge.json
Present when shared cases are enabled. Declares inputs and automationgauge-cases/
Optional Markdown cases for measurements you want to share or run in CIAGENTS.md
Guidance added by init --cases so an agent can find the cases
CAVEATS
This is not ThoughtWorks Gauge, the acceptance-testing tool. Packages named `gauge` in apt, Homebrew, or GitHub releases are usually that other project. This CLI is installed as `@withgauge/cli`.Local runs need a working Docker engine and agent credentials. Usage is billed by the agent provider, not by a Gauge package fee described in the CLI readme..gauge/ holds durable results. Deleting it drops the local history.
HISTORY
Gauge (withgauge) is published by gauge-sh as `@withgauge/cli`. The skill that teaches agents the CLI lives in the companion repository gauge-skills. Product documentation is at docs.withgauge.com.
