Walkthrough · with an AI agent
Asking an assistant about your data
The engineer wires Regixo into an AI assistant once — a minute of setup. After that a colleague who has never opened a terminal spends a sitting asking about the data in plain language, and the assistant answers from the real catalog. No commands, no exports, nothing leaves the machine. Here is that afternoon, question by question.
What this needs
Regixo already scanned (see Your first session), and an AI tool that
speaks MCP — Claude Desktop, Claude Code, Cursor, VS Code, or another. Screens below are static
examples of what you'll see, not live widgets.
1
Wire it once engineer
Open the portal, go to Help ▾ → Agents (MCP), pick the tab for your AI
tool, and press Copy snippet. Paste it into the file the page names for that tool and save.
(Prefer the terminal?
regixo mcp --print-config prints the same thing.) That is the
whole setup — the connection is read-only and stays on this machine.Connect your AI assistant to your data
local · stdio — runs inside your AI app1
Add Regixo to your AI tool
Pick your tool — we’ll show the exact snippet and the file it goes in (Claude Code takes a single command instead). Paste it and save. (regixo mcp --print-config prints this for you too.)
claude_desktop_config.jsonread-only
{
"mcpServers": {
"regixo": {
"command": "npx",
"args": [
"regixo",
"mcp"
],
"env": {
"REGIXO_DATA": "/Users/you/app/.regixo",
"REGIXO_CONFIG": "/Users/you/app/.regixo/regixo.yml"
}
}
}
} Paste into macOS:
~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json This pins your project’s data dir + config, so the agent reads this catalog from wherever your AI app launches it.
📘
Hand your agent the reading playbook. One command prints a short markdown skill that teaches your agent when to call which tool — and how to read the trust signals (draft vs confirmed, staleness) — before it answers. Paste it into your agent’s instructions (an AGENTS.md, or your tool’s rules file).
2
Restart the app, then ask it about your data
That’s it — your agent can now see the catalog. Try a question to confirm the connection; you’ll see it call Regixo’s tools before it answers.
You
“Where do we store customer email, and is any of it special-category data?”
A one-time paste. From here the engineer steps away — the rest needs no commands.
Tip: also run
regixo mcp --skill and paste the short playbook
into the agent's instructions, so it knows when to call which tool and how to read a draft.2
Hand it over colleague
The colleague opens their AI assistant and just types questions in plain English.
Behind each answer, the assistant calls Regixo's read tools — you'll see the calls listed before the
reply. The Agents page names every one of them, and the page's own summary is the guarantee: 8 read
tools · read-only · metadata only. There is no write tool to call, so the assistant can read the
map but can never flip a flag or confirm a legal field — and over stdio, nothing leaves the
machine.
What your agent can read 8 read toolsread-only · metadata only▾
search_datasetsFind tables by name, schema, column or PII kind — each hit with a why and a snippet.
get_datasetOne table’s columns, types and per-column personal-data flags, plus its lineage.
get_lineageHow data flows — the detected and asserted edges.
get_compliance_stateThe RoPA draft, each legal field’s status reported — never set.
get_provenanceThe trust signal — reachability, staleness, classifier method.
get_changesThe append-only log — what moved in your map, and when.
get_glossaryYour business terms, definitions and the datasets they link to.
get_doraThe DORA register draft (regulated financial firms) — read-only, like the rest.
prefer HTTP?The same catalog, read-only, over
/api/v1 on localhost:4319 — backed by the same reader, so it always agrees with the agent.The colleague types a question. The assistant reads the catalog and answers. That
is the entire loop for the rest of the sitting.
3
"Where do we store customer email?" colleague
The first question. The assistant searches, opens the best hit, and checks how
trustworthy the answer is before it commits to one.
what you'll see — inside your AI tool
Where do we store customer email?
✦ search_datasets q:"email"
✦ get_dataset id:"app-db/public/users"
✦ get_lineage id:"app-db/public/users"
✦ get_provenance
Customer email is stored in
app-db/public/users.email — flagged personal data, matched by the name-rule
email — and in stripe/customers, which mirrors it for billing.
⚠ One caveat: legacy-crm
hasn't been scanned since 12 Jun, so if email also lives there this list may be incomplete.
A real answer with a real limit attached. Because
get_provenance reported
a stale source, the assistant says so instead of pretending the map is complete.Same sitting, next question →
4
"Is any of it special-category?" colleague
A natural follow-up. Special-category data (Article 9 — health, beliefs, and so on)
carries stricter obligations, so the colleague wants to know where it is.
what you'll see — inside your AI tool
Is any of it special-category?
✦ get_dataset id:"app-db/public/health_survey"
Yes. In app-db/public/health_survey,
the column health_status is flagged ⚠ Art. 9 — special
category. Anything derived from it needs the stricter Article 9 handling; the rest of the email
data is ordinary personal data.
The Article 9 flag comes straight from the classifier's per-column labels — the
assistant reads it, it does not decide it.
5
"Is our RoPA up to date?" colleague
Now a compliance question. The assistant reads the record's draft state and
reports it honestly — what a machine suggested versus what a human has confirmed.
what you'll see — inside your AI tool
Is our RoPA up to date?
✦ get_compliance_state
It's a DRAFT — nothing here
is signed. Across your activities, these legal fields are still suggested
and need a human on the compliance team to confirm them:
• Lawful basis — suggested · Retention — suggested · Transfers outside the EU — suggested
✓ Data categories — found (a mechanical fact, read from your schema's personal-data flags — not a legal call)
I'm reporting the suggestions as
suggestions — I can't treat a suggested field as your settled position.
The assistant names which fields are still guesses and which a person has confirmed
— and never presents a suggested legal field as decided.
Why it can't just fill them in: see the honesty note below.
6
"What does controller mean, and does DORA apply to us?" colleague
Two quick checks — one on a term the record uses, one on whether the DORA register is
even in scope. The assistant reads the team's glossary and the DORA state; it never decides either.
what you'll see — inside your AI tool
What does controller mean, and does DORA apply to us?
✦ get_glossary term:"controller"
✦ get_dora
Controller, from your glossary —
the party that decides why and how personal data is processed (GDPR Art. 4(7)).
The DORA register isn't in scope
here — it's switched on only for regulated financial firms, and this catalog isn't marked as one.
If that changes,
get_dora returns the register.A definition read straight from the team's own glossary, and an honest "not in scope"
for DORA rather than an invented register.
7
"What changed this week?" colleague
The last question of the afternoon. The assistant reads the append-only change log and
summarises the newest entries.
what you'll see — inside your AI tool
What changed this week?
✦ get_changes
Three things moved since Monday:
• app-db/public/users — column
marketing_consent added• logs.ip_address — reclassified to not personal (by a person)
• lineage asserted: stripe/customers → users
Newest first. That new
marketing_consent column may touch a signed activity — worth a look.A plain summary of the diary the catalog keeps for you — the same feed a scheduled
re-scan appends to.
One sitting, no commands
The colleague now knows what to fix (a stale
legacy-crm, a fresh marketing_consent
column) and what to forward to the compliance team (the still-suggested legal fields) — without typing
a single command. The engineer wired it once; the answers came from the real catalog.Where "no commands" stops
The connected assistant reads and answers — that is all it can do over MCP. It cannot scan,
correct a flag, confirm or sign a legal field, unlock the paid record, or pay: the read-only
regixo mcp server ships no write tool, so there is no path for it to change
anything. A different surface can change things: a coding agent with a terminal runs the ordinary
regixo commands (scan, correct a flag, describe). The line neither surface crosses is the
legal one: confirming a lawful basis, signing, unlocking or paying are human acts. When the assistant
reports the RoPA, it reports a draft and names what still needs a person.
Full detail on the connection, the eight tools, and the trust vocabulary: AI agents & MCP →