A weekend ↓ architecture v1.0 · 2026-08-12

The staff rota engine: publish the schedule once, and be able to prove it

Availability collected before the deadline, rest-time and contract-hour breaches caught before you send it, and a dated record of every schedule you published

Setup: A weekend (~6-12 h) Running cost: $5-15/mo (software included) Time saved: saves ~1-2 h/week for the owner at 8-20 staff on a weekly rota — basis After setup: You run it — collects, checks, publishes and archives by itself; deciding who works which shift stays with you Data lives: Your server — Availability, contract hours, absences and every published version stay on your box. No vendor holds your staff roster, and the retention period is one you set rather than one you inherit.

Gristn8nDockerTimefold SolverRemindersReporting

Built for small businesses in general — not a sector list.

Before you start

Your own notice deadline written down before you build anything: five working days for variable part-time rosters in Belgium unless a generally binding sectoral agreement shortens it, 72 hours for a New York City retail employer with 20 or more staff, otherwise whatever your contracts and work rules already promise
A machine that stays on and is reachable over HTTPS, because staff enter their availability from their own phones
Contract hours, required skills or certificates and maximum consecutive days per person, current on the day you build the table and re-verified every time a contract changes
A real minimum-staffing rule per hour of the week: how many people, and which skills must be among them, or the coverage check has nothing to test against
Agreement with your team on how weekends and public holidays are shared, before any of it is encoded
A privacy notice covering what the rota system stores about staff and how long it is kept, and an availability form with no field for the reason

What breaks first FAILS SILENTLY

the staff table behind the checks: a contract change, a new starter or an expired certificate that never reached it. Wrong inputs produce a green light on a rota that breaks someone's hours. Re-verify contract hours and skills on every contract change, and run a standing review twice a year.

Version history

v1.0 · 2026-08-12 — First release. Grist as the table, n8n as the pre-publication check and publisher, with the constraint solver named as the 25-staff upgrade rather than the default, and the EU notice-period rules stated as the reason the archive exists.

Verdict first: the planner is not the valuable part. Any spreadsheet can draw a week. What a spreadsheet cannot do is stop you sending a rota with an illegal gap in it, and it cannot prove what you published and when. Those two things are the build. If you have fewer than five staff on fixed hours, skip this and keep your spreadsheet. Between roughly five and twenty-five people on shifts, this is a weekend that removes a recurring argument.

The architecture

the same flow as text

      [ staff ] → enter unavailability and holiday, one form

      [ you ]   → contract hours, skills, minimum staffing per hour

              ▼

      [ Grist — one table, on your own box ]

              ├─ deadline reminder, then a chase to whoever has not answered

              └─ you draft the week here, by hand, the way you always have

              ▼

      [ n8n — the check, before anything leaves the building ]

              ├─ less than 11 h between two shifts → named, with both shifts

              ├─ no 35 h break in the week → named

              ├─ rolling 48 h average, contract hours over or under

              ├─ an hour with nobody, or nobody holding the required skill

              └─ publishing later than your own notice deadline

              ▼  clean, or you fix it and run it again

      [ publish, once ]

              ├─ a dated PDF, filed and never edited

              ├─ a calendar feed per person, subscribed on their phone

              └─ one message to the team, timestamped

      Every change after publication writes a new dated version. Nothing overwrites.

Three parts, and the middle one is the product. A table anyone can edit, a check that runs between drafting and sending, and a publish step that leaves a record it cannot quietly change. You still decide who works Saturday. The machine collects, checks, sends and remembers.

The problem

The rota costs a small business two separate things, and only one of them is time. The time is the chasing: who is away in week 34, who swapped with whom, and the six messages that arrive after you send it. The other cost is that in the EU the moment you publish a schedule is a regulated moment, and most small teams cannot say what they published or when they published it.

What the rules actually ask for. Directive 2003/88/EC sets three numbers you have to plan against: 11 consecutive hours of rest in every 24 (Article 3), 24 uninterrupted hours plus that 11 in every seven-day period (Article 5), and a 48-hour average per seven days including overtime (Article 6b), averaged over a reference period of up to four months (Article 16b, longer by derogation or collective agreement). Directive (EU) 2019/1152 adds the part people miss: where the work pattern is mostly unpredictable, staff must be told the reference hours and days they can be asked to work and given a reasonable minimum notice, and if either is missing they may refuse the shift with no adverse consequences. It applies from an average of more than three hours a week over four weeks, so almost everyone with a weekend student is inside it.

In Belgium the notice period is a hard number. For part-timers on a variable schedule, the individual rosters must be communicated in a written, dated notice at least five working days in advance. A generally binding sectoral collective agreement can shorten that, but never below one working day. The notice, or a copy, has to sit where the work rules are consulted for as long as the roster is in force, and be kept for a year after it stops being in force. That is a document retention rule, and it is the reason a rota that only ever existed in a group chat is a problem.

In the US the same idea arrives as predictive scheduling, mostly aimed at chains. Oregon requires a written schedule 14 calendar days ahead and pays a premium for employer-made changes, but only for retail, hospitality and food employers with 500 or more workers worldwide. Seattle's ordinance also starts at 500. The one that reaches a genuinely small business is New York City: its Fair Workweek Law covers any retail business with 20 or more employees across all its NYC stores under common ownership, counting part-time and seasonal. Those employers must give the schedule 72 hours before the first shift, may not use on-call shifts, may not cancel or cut a shift by more than 15 minutes inside 72 hours, and must hand over any past schedule from the last three years within 14 days of a worker asking. First violation, $500.

What you would otherwise rent. Planday lists from £2,99 per user per month with a five-user minimum. Papershift's own booking page runs $4, $6 and $9 per employee per month excluding VAT. Shiftbase's tiers start around €30 a month including six employees and around €72 including twelve, plus €4 to €6 per employee above that. Per-location tools sit higher: an April 2026 comparison of restaurant tools puts Combo at about €60 a month per location and Skello at about €79. At ten staff that is roughly €350 to €900 a year against €60 to €180 for the box below. Be honest about that gap: a few hundred euro a year is a reason, not a compelling one. The compelling one is that the check and the archive are yours, and neither is a feature you can buy back cheaply once a dispute starts. List prices move, so verify before you quote them.

Tool choices — and why

Grist

(Apache-2.0, 11.441 stars, pushed 11/08/2026) holds the staff table, the availability and the draft rota. It is picked over a spreadsheet for one reason: it keeps a document history, so when someone says the rota changed after it went out, there is an answer instead of a memory. Formulas are Python, which matters when you want a column that computes hours per person per week and turns red at the contract limit. If you already run NocoDB (64.498 stars, pushed 11/08/2026) for something else, it does the job too. The shape of the table matters more than the brand.

n8n

(Sustainable Use Licence, fair-code, 200.261 stars, pushed 11/08/2026) is the check and the publisher. Four small workflows: the availability deadline and its chase, the pre-publication check, the publish step that renders the PDF and the per-person calendar files, and the change log. If you already run n8n for the late-invoice chaser or the no-show reminders elsewhere in this library, this is more workflows on the same box. Read the licence once if you intend to sell this setup to clients, because that is where its terms bite.

How staff see it

A calendar subscription per person, one secret URL each, added once to the phone calendar and never thought about again. That kills most of the "am I working Thursday" traffic on its own. If you want proper two-way calendar accounts rather than read-only feeds, Radicale (GPL-3.0, 4.895 stars, pushed 08/08/2026) is a small CalDAV server that will do it. For the "the rota is out" push, ntfy (Apache-2.0, 33.401 stars, pushed 11/08/2026) sends a phone notification without an SMS bill.

The optimiser, and when it is worth it

Timefold Solver (Apache-2.0, 1.669 stars, pushed 01/06/2026) is the serious open-source constraint solver for exactly this problem, and its employee-scheduling quickstart (Apache-2.0, 554 stars, pushed 12/06/2026) is a working Java service that assigns shifts against availability and skills. It is genuinely good, and for eight people it is the wrong tool. Writing the constraint model honestly takes longer than a year of Saturday mornings drawing the rota yourself, and a solver that produces a mathematically fair week your staff experience as unfair costs you more than it saves. The threshold where it starts paying: roughly 25 or more people, several skills that are not interchangeable, more than one site, or a rota you rebuild more than twice a week. Below that, run the check and draft by hand.

What we did not pick

No language model, anywhere. Whether two shifts are 11 hours apart is subtraction, and a model that is right 97 % of the time on a compliance check is worse than useless, because the 3 % arrives with a confident sentence attached. We also did not pick a full HR system. Frappe HR (GPL-3.0, 8.553 stars, pushed 11/08/2026) has shift management and would work, but it brings payroll, recruitment and onboarding with it. Take it if you want the whole HR platform; do not take it to solve a rota. And we did not pick the group chat, which is the real incumbent: it has no deadline, no version, no record, and it puts the burden of remembering on the one person who cannot afford to be wrong.

Monthly cost

Software €0 for everything above. Hosting: Grist, n8n and a database fit on one small VPS at roughly €5 to €15 a month, and Timefold, if you ever add it, wants about 2 GB of RAM to itself. Notifications are free over calendar feeds and ntfy. SMS is the only line that costs per message, and at rota volume you do not need it.

Setup outline

1. Write down your own notice deadline first, before you touch any software. Belgium: five working days unless a sectoral agreement in your joint committee says otherwise. NYC retail at 20 or more staff: 72 hours. Elsewhere: whatever your contracts and work rules already promise. Everything below is built backwards from that date.
2. Build the staff table in Grist: name, contract hours per week, skills or certificates that gate a shift, maximum consecutive days, and any fixed unavailability. This table is the one that goes stale, so put a review date on it now.
3. Add the availability form and a deadline two days before you draft. Ask when someone cannot work. Do not ask why. What you will see when it works: a table that fills itself in over the week instead of arriving as messages.
4. Wire the chase in n8n: at the deadline, a single message to whoever has not answered, and a summary to you. One chase, not three.
5. Draft the week by hand in the rota table. You are not automating this part yet, and possibly ever.
6. Build the check as one workflow that reads the draft and returns a list, not a verdict: rest under 11 hours between consecutive shifts, no 35-hour break in the seven-day period, rolling 48-hour average against your reference period, hours over or under contract per person, hours with nobody rostered, hours with nobody holding a required skill, and whether the publish date is inside your notice deadline. Every finding names the person and the two shifts that caused it. What you will see when it works: a green line, or a list you can act on in two minutes.
7. Build the publish step: render a dated PDF and file it read-only, write one calendar file per person to a secret URL, send one notification. Then test the reverse: pull last month's PDF out of the archive and check it is the version people actually worked.
8. Add the change log. Any edit after publication produces a new dated version and a message naming what moved and how much notice the affected person got. Run the whole thing in parallel with your current method for four weeks before you rely on it.

Pitfalls — the real ones

Publishing is a dated act, not a convenience

In Belgium the notice with the individual rosters must be written, dated, communicated at least five working days ahead, kept where the work rules are consulted while it is in force, and retained for a year after it stops being in force. Deviations from the communicated schedule get recorded separately and kept for five years. If your rota lives only in a chat thread, none of that exists.

The clopening is illegal before it is unpopular

Closing at 23:00 and opening at 07:00 is eight hours apart, and Article 3 wants 11. This is the single most common finding when the check first runs, and it is usually the same two people every month. Oregon prices the same idea at time-and-a-half inside 10 hours; NYC fast food charges the employer a $100 premium per clopening. Different mechanisms, one message.

An average is not a week

The 48-hour limit is an average over a reference period of up to four months, so one 52-hour week is not automatically a breach and a checker that flags it as one will be ignored inside a fortnight. Make the check say which rule it is testing and over what window, or it becomes noise.

The green light is only as good as the table

A contract change that never reached Grist produces a clean check on a rota that breaks someone's hours. This is the failure mode that does not announce itself. Re-verify contract hours and skills whenever a contract changes, and put a standing review in the diary twice a year.

Collect availability, not reasons

"I cannot work Tuesday evening" is scheduling data. "I have dialysis on Tuesday evening" is health data, and once it is in your table it is a category you now have to justify holding. Build the form so the reason field does not exist.

Staff data on your box is still staff data

A rota system holds names, hours, absences and sometimes certificates. Tell people what it stores and how long you keep it, keep it to the retention the law asks for rather than forever, and do not quietly extend it into monitoring who is late.

Do not let a solver own fairness

If you do reach the size where Timefold earns its place, weekends and public holidays are the constraint people feel, and an optimum that spreads them evenly on paper can still land badly. Fix the fairness rules with your team first, then encode them.

One publication, one channel

The failure is not a missing rota, it is two rotas. If the PDF, the calendar feed and the chat message can disagree, someone will work the wrong version. Publish all three from the same step or drop one.

Verified repos

Grist — Apache-2.0, 11.441 stars, pushed 11/08/2026
n8n — Sustainable Use Licence (fair-code, free for internal business use), 200.261 stars, pushed 11/08/2026
Timefold Solver — Apache-2.0, 1.669 stars, pushed 01/06/2026
Timefold Quickstarts (employee-scheduling) — Apache-2.0, 554 stars, pushed 12/06/2026
NocoDB — 64.498 stars, pushed 11/08/2026
Radicale — GPL-3.0, 4.895 stars, pushed 08/08/2026
ntfy — Apache-2.0, 33.401 stars, pushed 11/08/2026
Frappe HR — GPL-3.0, 8.553 stars, pushed 11/08/2026

Get the next blueprint as it publishes — free, practical, verified tooling only.

Subscribe · Want this audited or extended for your business? DM @RXed_EU