open source

Agent tools

understudy-agent-tools is the MIT-licensed local client: a library of skills that teach a coding agent the Understudy methodology, plus a thin CLI for the durable parts — auth, artifact checks, monitored runs. It works offline; the hosted gateway is something you opt into.

The repository

github.com/understudylabs/understudy-agent-tools — public, MIT. Two layers with a deliberate division of labor:

layer	what lives there
skills/	Progressive-disclosure playbooks for coding agents (Claude Code, Cursor, Codex): capture evidence, understand a workload, run evals, optimize, compare local models, hand off to training. The judgment lives here.
src/ (the CLI)	Thin TypeScript: login, doctor, workload/route/capture shortcuts, durable runs an agent can monitor. Deliberately boring — reliable execution, auth injection, and safety gates only.

The workflow the skills encode is local-first:

capture evidence -> attach harness/environment
  -> confirm metric/validator/holdout -> rerun baseline
  -> optimize workload -> conservative claim packet

No registration is required for that loop — fixtures and local captures work offline. understudy login adds hosted gateway access when you want live traffic, managed models, and the dashboard, at which point the CLI drives the same control plane API documented on this site.

Install

fast first run

curl -fsSL https://raw.githubusercontent.com/UnderstudyLabs/understudy-agent-tools/main/install.sh | bash

The installer sets up the CLI, installs the Claude Code plugin when claude is available, and opens Claude Code in the current directory — then /reload-plugins and /understudy:onboardtake over. It's resumable (--resume), supports non-interactive runs (--yes), and deliberately does not download model weights or make frontier calls — those happen inside the skill flow, with consent.

as a Claude Code plugin (from a clone)

claude plugin marketplace add /path/to/understudy-agent-tools
claude plugin install understudy@understudy-skills
# in your session: /reload-plugins, then /understudy:onboard

Installing the plugin is the recommended way in: the skills are what let an agent explain Understudy and walk you from a trace to a shipped improvement. Removal is symmetric (claude plugin uninstall) and touches nothing outside the plugin registry.

Default posture

No provider calls, uploads, model downloads, or hosted jobs run by default — local-first means local until you say otherwise.
After authentication the CLI emits bounded product telemetry (documented in the repo's docs/telemetry.md); disable with UNDERSTUDY_TELEMETRY=0.
Secret values are never inspected or echoed.