understudy

Understudy is a public, MIT-licensed skill library plus a thin CLI. It gives your coding agent the playbooks to evaluate an AI workload locally, optimize it, and route it to a better-value model — local-first, with no uploads, no provider calls, and nothing spent by default.

Get started (30 seconds)

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

The installer sets up the CLI, asks which coding agents to attach to (Claude Code, Cursor, Codex, OpenCode, Hermes Agent, all, or CLI-only), and — if you pick Claude Code — opens it in the current directory. Then, in Claude Code:

/reload-plugins
/understudy:onboard

That's it. Onboarding profiles your machine, interviews you about your workload, and starts the improvement loop. No registration required. On other agents the activation command differs slightly — see Pick your platform.

The installer intentionally does not download model weights, start MLX, launch the ladder server, or make frontier calls. Those happen inside the skill flow, where the agent explains tradeoffs and asks consent first.

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

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

curl -fsSL https://raw.githubusercontent.com/UnderstudyLabs/understudy-agent-tools/main/install.sh | bash -s -- --from-step 2

npm install
npm run build
node dist/bin.js --help

npm install -g @understudylabs/understudy-agent-tools
understudy spine

What it does

• Captures evidence from your real LLM traffic — traces, transcripts, or a public benchmark — into a local, redacted evidence base.
• Builds a real eval — harness, environment, metric, validator, and holdout — so improvements are measured, not vibed.
• Optimizes the workload — prompt optimization (GEPA/DSPy), local model labs, distillation, and RLM training, all runnable on your machine.
• Picks a route — compare model sweeps, estimate hosted costs, then ramp a percentage of traffic to a cheaper model while the rest stays on your current provider.
• Ships a conservative claim packet — what changed, how it was measured, and what it saves, stated only as strongly as the evidence supports.

The whole OSS loop runs locally and is skill-led:

flowchart LR
    A["Capture<br/>evidence"] --> B["Attach harness +<br/>environment"]
    B --> C["Confirm metric,<br/>validator, holdout"]
    C --> D["Rerun<br/>baseline"]
    D --> E["Optimize<br/>workload"]
    E --> F["Conservative<br/>claim packet"]

Registration is not required for that loop. Hosted gateway access is available after understudy login; browser, channel, daemon, and desktop-runtime commands remain outside this public CLI until intentionally extracted.

How it's built

Understudy is agent-platform-neutral at the skill layer. All six supported coding agents share the same skills/ tree and differ only in a thin adapter: manifest, install path, reload step, and onboarding invocation. The adapter registry lives in docs/agent-platform-adapters.md and is exposed through understudy platforms.