Your pilot: day 0 to live

Welcome — if you’re reading this, your pilot just started. This page is the whole journey, in order: what you already have, what happens on the kickoff call, and what “live” means. Your onboarding engineer follows exactly this track (it’s literally a runbook they run), so nothing that happens will surprise you.

A pilot is 90 days of the full Scale surface — every agent, governed writes, the context graph, the Spellbook Data Catalog (research preview), hosted Conductor, and a named engineer — free, no card. The terms you signed also fixed the exit up front: what you keep and what ends was agreed before anything installed.

Before day 0 — what you should already have

From your welcome email:

• Your customer npm token — scoped, read-only, expiring on your pilot end date. This unlocks the paid packages; nothing else installs without it.
• Your kickoff call invite (30–60 minutes) with your named onboarding engineer.
• A link to this site.

What your side should line up before the call — one engineer’s laptop with Node.js 20+ and their usual coding agent (Claude Code, Cursor, Codex CLI, or OpenCode), read credentials for the first systems (created per least-privilege guidance), and a decision on the first schema that matters — the scoped slice you’ll go live on. If your IT or security team needs to pre-approve anything, hand them the procurement & security package — it lists every prerequisite and everything that does and doesn’t leave your network.

Checkpoint: token in hand, call booked, one schema chosen.

Day 0 — the kickoff call

Your onboarding engineer runs the onboard-data-workers skill with you, live. In one call:

1. Install — the agent swarm lands in your engineer’s coding agent (claude mcp add data-workers -- npx -y dw-claw, or the equivalent for your client), and answers questions on sample data before any credential exists. That’s the 🟡 Evaluation state — deliberate, so you evaluate before you connect.
2. Connect — your engineer sets the environment variables for the in-scope systems (exact names per system). Credentials stay on your machines.
3. Verify — every connection gets a live test, on the call. Nothing is called Connected until its test passes — the three-state model is the product’s core promise, and your onboarding models it.

Checkpoint: every in-scope system shows 🟢 from a test you watched run.

Day 0, continued — context, console, Conductor

4. Warm the context graph — your scoped schema is harvested into the Data Context Wizard, spot-checked against what your team knows is there, and your data owner promotes the first authoritative facts. (How warming works.)
5. Stand up the console — the Spellbook Data Catalog (research preview) runs locally with your npm token; connectors get registered and Test connection-verified in the browser.
6. Choose your Conductor mode — Approve-edits first, always; flip to Auto later in one click. Irreversible actions ask first in every mode, enforced in code. (Conductor.)

Steps 5 and 6 are their own decision-making leg — the no-touch list, approver naming, and the graduation path from Approve-edits to Auto are written up in Spellbook & autonomy guardrails.

Day 0, last step — your eval baseline

The eval-benchmark skill builds an eval set from your team’s real historical queries, freezes today’s correct answers from your systems of record, and scores the agents. That day-0 number is what the rest of the pilot is measured against — re-run every two weeks, the trendline is the evidence your pilot decision reads from. Failures aren’t hidden; they’re the hill-climb queue we work during your pilot.

Checkpoint: eval set committed in your repo, baseline scored, onboarding report delivered.

Weeks 1–12 — expand, measure, decide

• Expand scope on green: one schema at a time, one team at a time — each new connection earns its 🟢 the same way.
• Cadence: a short check-in roughly every two weeks, with a usage report built from real counters and your re-run benchmark score — never extrapolated.
• Anything broken or missing goes through the feedback pipeline (your engineers or their AI agents can file directly) — connector priority is demand-driven, and your pilot’s filings jump the queue.
• Two weeks before day 90: the conversion conversation, against your own evidence pack. Convert to Scale or Enterprise, or wind down cleanly — your data was never in our hands either way.

Getting help during your pilot

You have three routes, fastest first: your named onboarding engineer (the person from the kickoff call — your default for anything blocking), your pilot support channel (set up at provisioning), and the feedback pipeline for bugs, requests, and incidents — filed by your engineers or their AI agents, with incident filings routed ahead of everything else. If an internal owner on your side is the blocker (IdP admin, proxy owner), say so at the check-in — unblocking sequencing is part of our job, not a complaint.

If your deal is Enterprise-shaped

SSO/SAML, SCIM, dedicated VPC or your cloud, PrivateLink, on-prem or air-gapped — those are deployment add-ons your security org asked for, and they’re set up with us, not self-service. The Enterprise deployment track covers what happens and in what order; the day-0 journey above stays the same.