Walkthrough · the engineer

Running Regixo by talking

One sitting, start to finish, without typing a regixo command: an empty project, a coding agent, and a forwarded compliance draft at the end. Every step shows what you say, what your agent does, and — the part that matters — how you check it did it right.

What you need A coding agent open in your project — Claude Code, Cursor, Codex, VS Code, or any tool that can run a terminal command. Node 22.18+, and a database you can reach. There is nothing to install and nothing to register: your agent runs the same regixo command you would, in the same shell. See Work with your agent.

What you will not find below No invented agent replies. We do not know what words your tool will choose, and quoting a made-up one would teach you to expect a script instead of reading the result. Every step states the outcome Regixo produces — which we do know — and the signal you check it against. Portal screens are static examples of what you’ll see; they look the same whether you ran the scan or your agent did.

Ask for a data map engineer

You have a Postgres behind DATABASE_URL and a Stripe account, and no Regixo yet. You do not tell your agent which commands to run — you tell it what you want.

say

“Set Regixo up in this project — scan my sources and build the data map.”

Show the commandHide the commandShow the sentenceHide the sentence

run

$ npx regixo start

then

It reads Regixo’s manifest first (regixo help --json — every command, flag and error code, as data), so it is not guessing. Then it finds the sources it can already reach from your .env and scans names and types only. Regixo itself checks your Node version, before anything runs.

One question is yours, not its: is your organisation a regulated financial entity? Regixo cannot infer that, and it decides whether a DORA register is drafted — so it is a declaration, never a guess.

Check it worked: it should report a data map, a dataset count, a count of columns that look like personal data, and a DRAFT GDPR record. Those numbers — not a claim that it worked. Step 3 is how you confirm them.

Show what it prints in the terminalHide the terminal outputShow what your agent reportsHide what your agent reports

what you'll see — terminal

Regixo — free to install and use. It maps your data and drafts your compliance record, right here. Nothing leaves this machine.

▸ Looking for your databases…
✓ Found your Postgres database — ready to map  (from the DATABASE_URL setting)

  ? Connect & scan this source? [Y/n] y
  ? Is your organisation a regulated financial entity (a bank, payment/e-money institution, investment firm, or insurer)? [y/N] n

▸ scanning your sources — names and types only, never the values inside…

✓ Your data map and draft record are ready.
  To see them, type:  regixo open   (opens them in your web browser)

A map of 63 datasets and a first GDPR record draft — both built on your machine, from names and types. No row value was read.

Next: the Stripe key it could not find.

It stops at the secret — and that is correct engineer

Stripe needs an API key, and your agent will not ask you for it. It fills in every non-secret part, leaves the token’s slot marked, and hands back.

say

“Add our Stripe account to Regixo.”

Your agent fills in everything except the token — you put that in .env yourself.

Show the commandHide the commandShow the sentenceHide the sentence

run

$ regixo add stripe --ref STRIPE_API_KEY

then

Regixo writes <PASTE-YOUR-TOKEN-HERE> into .env for you, and the config keeps only the variable’s name — never the value. Your agent tells you which line needs you, and why it did not ask: a secret pasted into a chat is sent to the model and kept in the transcript on disk.

This is the agent working, not failing. Paste the key into .env yourself, then ask it to carry on. Until you do, Regixo answers SECRET_PLACEHOLDER_UNREPLACED — a checkpoint, not an error.

Show what it prints in the terminalHide the terminal outputShow what your agent reportsHide what your agent reports

example output — .env

STRIPE_API_KEY=<PASTE-YOUR-TOKEN-HERE>

Then regixo test stripe to confirm the key works — it connects and reads nothing.

The one thing only you can do, done by you — and nothing secret in the transcript.

Next: check what you actually got.

Check it — against Regixo, not against the agent engineer

This is your new job. An agent’s own summary of its work is not evidence — the coverage line is.

say

“Summarise what Regixo has connected and scanned, and what’s still missing.”

Show the commandHide the commandShow the sentenceHide the sentence

run

$ regixo status

then

Read the coverage line. A source that was not reached is missing from the map — and this is the failure that hides, because the counts still print, they are just quietly too low. So Regixo does not make you infer it: the bad case names the source.

If a source dropped out: ask your agent to run regixo doctor. It probes each source and names the exact fix. A database on a network your agent isn’t on will never connect from here — that is not the agent’s fault, and the three honest routes are in Connect your sources.

Show what it prints in the terminalHide the terminal outputShow what to checkHide what to check

example output

Regixo status — /Users/you/app/.regixo
  sources:  2 configured · 2 reachable
  datasets: 63 mapped · 55 hold personal data
  columns:  398 in total · 170 hold personal data
  coverage: ✓ all 2 source(s) reached
  record:   draft RoPA tracks your map — open it anytime: regixo open

Had Stripe failed, the last line would read coverage: ⚠ 1/2 sources reached — the map EXCLUDES stripe (fix: regixo doctor).

Two sources reached, and a count you recognise. That is the whole check — and it took one sentence.

Next: look at what it built.

Read the map yourself

The dashboard is the one surface that is the same either way — it reads what the scan produced, whoever ran it.

say

“Open the Regixo dashboard in my browser.”

Show the commandHide the commandShow the sentenceHide the sentence

run

$ regixo open

what you'll see — the Map's toolbar (the coverage meter) and one source fold

Regixo data catalog · free & local

Data map

63 datasets · 170 personal-data columns

10 of 10 sources reached

The coverage meter is the same signal as the ✓ list, drawn: every source reached. If one had dropped out the meter would say so — and that whole source would be missing from everything downstream.

Next: give the map some meaning.

Let it draft, and keep the confirming for yourself engineer

A scan captures structure. It cannot say what a table is for. Start by having Regixo draft that from the metadata — one command covers every dataset and every column, free and instantly. Everything lands as a suggestion, and stays one until you say otherwise. (Your agent can then rewrite the ones that read badly in your own words — that is the step after this.)

say

“Have Regixo draft descriptions for every dataset and column in my catalog from the scan — no model, just the metadata.”

Show the commandHide the commandShow the sentenceHide the sentence

run

$ regixo describe draft --all

then

Regixo composes these itself, from metadata only — humanised names, the personal-data kinds present, lineage neighbours. No model is involved and nothing is sent anywhere. Each one lands labelled suggested.

If your agent writes one in its own words instead (regixo describe set <id> "…" --model), that one renders as “suggested · AI-drafted” — visibly distinct from a human’s. Either way: a suggestion.

Check it worked — and then do your half: read them on the Map. They say suggested, and they will keep saying it until you act. Confirm is a person’s act — it records who signed off — so it carries no sentence and your agent will not run it. Same for a glossary term.

Show what it prints in the terminalHide the terminal outputShow what your agent reportsHide what your agent reports

example output

Drafted 12 suggested description(s). Review: regixo describe list

Twelve datasets that mean something now — and not one of them claims to be confirmed.

Next: the fields no agent may touch.

The legal calls it will not make

Regixo drafted a GDPR Article 30 record from the same scan. Its purpose, lawful basis and retention fields are blank on purpose.

run

$ regixo annotate set <activityKey> lawfulBasis "Art. 6(1)(b) contract" --confirm

This one’s yours. Only a person can confirm a legal field.

No sentence sits beside that command anywhere in these docs — because offering one would invite a machine to do the single thing it must never do. Your agent is told to stop and say why if you ask it to confirm.

Next: hand the draft to the people whose call it is.

Forward it — with your word, not its assumption engineer

The legal calls belong to your compliance team. Forwarding the draft is the last thing you do.

say

“Forward my Regixo draft record to our compliance team. Show me what would be uploaded before you send it.”

Show the commandHide the commandShow the sentenceHide the sentence

run

$ regixo invite

then

This one asks before it sends. regixo invite previews exactly what leaves the machine — table and column names, types, owners; never a row value — and waits. The playbook tells your agent to show you that preview and get your word in advance, never to assume it.

Check it worked: you get a DRAFT PDF you can email and a claim link your compliance team opens with no account. If your agent sent it without showing you the preview first, that is a real problem — say so.

Show what it prints in the terminalHide the terminal outputShow what your agent reportsHide what your agent reports

what you'll see — terminal

▸ this will upload a metadata-only snapshot. What leaves:
  · table & column names, types, owners      ✓ leaves
  · your row values                          ✗ never
  ? Send this snapshot?                        [y/N] y
✓ RoPA_DRAFT.pdf written — forward it to your compliance team
✓ claim link: https://app.regixo.com/claim/clm_7Q…

A PDF you can email and a link they open with no account — the hand-off is done, and you never typed a command.

What you did, and what you didn’t You said seven sentences. Your agent ran the commands, read the output, and stopped twice — once at a secret, once at a legal field. Both stops were the product working. Everything you checked, you checked against something Regixo said, not against something the agent claimed. That is the whole method: Work with your agent.

← Also

Your first session (at the terminal)

The hand-off round-trip