🔧 For the engineer·Lived walkthrough·see the whole journey ↗

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, not live widgets — and the assistant’s replies are not scripted here. Each one shows the question, the tools the assistant calls, and the catalog facts those tools return; the words your assistant wraps them in are its own.

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.

what you'll see — Agents (MCP) · the two setup steps: pick your tool, copy the snippet

Regixo data catalog · free & local

Connect your AI assistant to your data

local · stdio — runs inside your AI app

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).

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.

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 you'll see — Agents (MCP) · the eight read tools (close-up)

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.

"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.

the question, the tools called, and what they return

Where do we store customer email?

✦ search_datasets q:"email" ✦ get_dataset id:"app-db/public/users" ✦ get_lineage datasetId:"app-db/public/users" ✦ get_provenance

search_datasets → app-db/public/users · stripe/customers

get_dataset → column users.email — flagged personal data, matched by the name-rule email

get_lineage → app-db/public/users → stripe/customers (a billing mirror)

get_provenance → ⚠ legacy-crm last scanned 12 Jun — stale

Whatever words your assistant chooses, the substance is those four returns — and a faithful answer carries the last one as a caveat: a stale source means the list may be incomplete. The Map’s trust panel shows the same staleness signal, so you can check it yourself.

Same sitting, next question →

"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.

the question, the tools called, and what they return

Is any of it special-category?

✦ get_dataset id:"app-db/public/health_survey"

get_dataset → column health_status — flagged ⚠ Art. 9 special category; the dataset’s other flags are 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. The same flag sits on the Map behind the Art. 9 filter pill, if you want to see it with your own eyes.

"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.

the question, the tool called, and what it returns

Is our RoPA up to date?

✦ get_compliance_state

get_compliance_state → stamp: DRAFT — nothing signed

• 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)

Every field’s status is reported, never set — there is no tool that can confirm one, so a faithful assistant relays the suggestions as suggestions rather than treating a suggested legal field as your settled position. The Record page shows the same statuses, badge for badge.

Why it can't just fill them in: see the honesty note below.

"What does controller mean, and does DORA apply to us?" colleague

Two quick checks — one on a term your team has defined, one on whether the DORA register is even in scope. There is no "look up one term" call: the glossary comes back whole and the assistant picks the entry out of it. The assistant reads the team's glossary and the DORA state; it never decides either.

the question, the tools called, and what they return

What does controller mean, and does DORA apply to us?

✦ get_glossary ✦ get_dora

get_glossary → every term the team has defined; among them Controller — “the party that decides why and how personal data is processed (GDPR Art. 4(7))”

get_dora → not in scope — the register is generated only when regixo.yml: dora: true

A definition read straight from the team's own glossary, and an honest "not in scope" for DORA rather than an invented register.

"What changed this week?" colleague

The last question of the afternoon. The assistant reads the append-only change log and summarises the newest entries.

the question, the tool called, and what it returns

What changed this week?

✦ get_changes

get_changes → three entries, newest first:

• app-db/public/users — column marketing_consent added

• logs.ip_address — reclassified to not personal (by a person)

• lineage asserted: stripe/customers → users

A plain readout of the diary the catalog keeps for you — the same feed a scheduled re-scan appends to, and the same entries the What changed page shows. A new column like marketing_consent is the kind of change that can touch a signed activity — the walkthrough A change & a re-sign picks up exactly there.

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 →

Your next step That fresh marketing_consent column is the kind of change that can unsettle a signed record. See that drift caught end to end and re-signed by a person: A change & a re-sign →.

← Previous walkthrough

The hand-off round-trip

Next walkthrough →

A change & a re-sign