Skip to content
bucker

The day-one stack plan

npx create-bucker --stack reads your repository and produces one plan for the whole app instead of one SDK: the SDK and what it needs, a monitor for every scheduled job it can read without guessing, an HTTPS check on your production URL, a draft status page, and a starter alert to the two people you name. It tells you what the plan will cost a month before anything is applied, and it ends in a signed receipt with one line per signal.

# What a coding agent runs: prints the plan and stops. No DSN, no token, no network.
npx create-bucker --stack --json --dry-run \
  --url https://app.example.com --notify you@example.com --notify them@example.com

# What you run, after reading it:
npx create-bucker --stack --token "$BUCKER_TOKEN" --org acme --project shop \
  --notify you@example.com --notify them@example.com \
  --approve <the plan's bucker.batchDigest>

What it finds

Source Read from Becomes
Vercel Cron vercel.json crons a cron monitor (UTC)
GitHub Actions .github/workflows/*.yml on: schedule: a cron monitor (UTC)
node-cron cron.schedule('…') in a file that imports node-cron a cron monitor
BullMQ pattern / every in upsertJobScheduler or repeat a cron or interval monitor

It never invents a schedule. A job is found, not monitored, with the reason printed, when its expression is a six-field cron (a seconds field), a macro such as @daily, a variable or a template instead of a string, or when its timezone is not a literal; and a schedule() call in a file that does not use node-cron is not read as a cron at all.

The production URL comes from --url or package.json homepage, and only when it is https on a public host. A loopback, private-network, bare-IP or .internal name is refused with the reason, because the check runs from Bucker's address towards that host. The plan keeps the origin and path and drops any query string.

What it costs

The plan prices itself from the weights register, in the same units your usage page shows. HTTP-class and cron checks are unmetered (weight 0, never upsold). A cron check-in is an internal request: 0 units in the ledger, and the one-unit floor at the accept decision, so the plan states both numbers. The SDK's own volume depends on your traffic and is shown as such until you pass --errors-per-day; one error is one unit.

Two halves, reviewed separately

  • The repository half is the installer you already know (create-bucker's plan, with a diff for every file and a stop on a dirty tree). You commit it.
  • The Bucker half is a list of creations in your account (cron monitors, the URL check, the e-mail channels for the two people, the alert rule, the draft status page) under one digest. You approve it once: --approve <digest> binds exactly what you read (a plan that changed since has a different digest and is refused), or answer the prompt at a terminal. --yes does not approve it: that flag is about edits to your own code. Re-running adopts what already exists.

Publishing the status page is never a step: it is yours to do, from the dashboard. Deleting anything is never a step either; the receipt lists what the plan created so you can remove it.

The Vercel half (drains, deploy webhooks and BUCKER_DSN in the Vercel environment) waits for the Vercel marketplace integration and is listed in the plan as not available.

The receipt

After applying, the tool reports which signals the plan named and the server decides what each one's status is from what exists in your project:

Status Meaning
installed the object exists (the monitor, the check, the page, the rule, or the SDK receipt)
first_event_seen it has produced something: a check-in, a probe result, a real issue
not_seen not there yet, or there and silent
skipped the plan found it and did not plan it (a six-field cron, no URL, nobody to alert)

The server signs one stack_receipt claim into your organization's receipt log. It carries ids, statuses and times, never a path, URL, schedule or address, and bucker-verify verify-log checks it offline. A cron that has not ticked yet reads installed; run create-bucker --stack --receipt-only after it has to file a fresh receipt (the old one stays in the log). Read the newest with GET /orgs/:org/projects/:project/stack-receipt.

What the coding-agent run never sees

--json --dry-run makes no network call and needs neither a DSN nor a token; every line it prints passes the output guard that redacts DSN- and token-shaped strings, and the addresses you name appear masked (a***@example.com). The DSN is written to your env file by the installer and never printed.