Skip to main content
This first session verifies setup and creates an onboarding receipt. It stops before a model smoke, probe, or rules evaluation.

​
Stet is driven by your coding agent

You do not operate Stet directly. Your coding agent runs the CLI, and the skill aligns how it does so with your interests: propose a bounded plan first, ask before anything that spends model tokens, never quietly broaden scope to get a better-looking number, and report the canonical result honestly — including what it does not support. You set the question, approve the spend, and make the decision. The prompts in these docs are the words you say to your agent to start each bounded workflow. Evaluations spend model tokens through the provider account your agent already uses, so they draw on your existing usage limits rather than a separate bill. A full eval can consume a substantial portion of a $100 or $200 monthly subscription tier’s limits, which is why every workflow here states its expected spend and waits for your approval before launch. This quickstart itself spends nothing on models.

​
1. Install the CLI

macOS and Linux:
Windows PowerShell:
The installer scripts and agent skill come from the current public Stet-AI/stet-cli/main source. The installers select the latest stable CLI by default; verify the installed binary with stet --version. The CLI ships for Windows, but the documented end-to-end Docker-backed evaluation workflow currently covers macOS and Linux.

​
2. Sign in and install the skill

Commands that launch AI evaluation or regrading require sign-in; onboarding and reading results do not.
stet auth login opens a browser sign-in and stores a token on your machine; signing in for the first time creates your account and starts a trial automatically, with no payment details collected. You only need it for commands that launch new AI evaluation or regrading work — stet eval, stet monitor run, stet baseline rerun, and the stet runs repair/regrade family. Onboarding a repository, building your task corpus, and reading or gating existing results all work without signing in. Stet is local-first: your repository, patches, and evaluation artifacts stay on your machine. Evaluations bill model tokens to your own provider account, never to Stet; run stet auth status to see your account and trial expiry. Install the agent skill separately:
The CLI runs evaluations and manages artifacts. The skill tells your coding agent how to choose a workflow and read the canonical result. skills is the open skill-manager CLI: --skill stet selects the Stet skill from this repository, --all installs it for each supported coding agent it detects, and npx skills list shows what was installed and where.

​
3. Verify prerequisites

Run the checks that match the backend you plan to use:
The default isolated path needs Docker, Python 3.12 or newer, uv, and authentication for the model provider you eventually plan to evaluate. Python and uv power Stet’s replay harness regardless of your repository’s language; your repository itself does not need to use Python. GitHub CLI access is needed for private release overrides or PR-backed discovery.

​
4. Ask your agent to onboard one repository

From the repository you want to measure, say:
The installed skill carries the guardrails — a narrow test command, no quiet scope broadening, and a hard stop before model spend. If you want every guardrail explicit in the prompt itself, use the long form:
Your agent should inspect CI and build files, ask about the product areas and work mix you care about, select real merged work, and create the repository’s Stet harness files. A starter slice should cover the work you intend to track; it is not automatically representative just because it was recent. If Claude is the selected provider, use the secure setup-token path: run claude setup-token, store the printed token in ~/.config/stet/claude-oauth-token with 0600 permissions, and do not export the token into shell history or an environment variable. Stet reads that file.

​
5. Inspect the onboarding receipt

Ask your agent to summarize the receipt before doing more work:
If setup or task selection is blocked, use Troubleshooting. When the receipt is credible, continue to Your first eval. Do not treat this setup check as a model result. To expand a credible starter slice, ask for a bounded follow-up: