DKAN+AI: From Catalog to Conversation

AI agents on DKAN, connected with MCP (Model Context Protocol).

Dan Goodwin · Senior Engineer, Drupal + React · CivicActions

Modules: dkan_ai_query  ·  dkan_mcp_server

Dan Goodwin

Senior Engineer, Drupal + React
—
Apex, NC

• Drupal since D6; now focused on federal open data
• I write custom DKAN AI modules and the React frontends that consume them
• Exploring how AI can lower the barrier between open data and the people who need it

Off the clock: guitar, Aikido.

Open data is a simple promise

Governments and institutions publish what they know — and let anyone use it.

• In the U.S., it's often the law. Federal agencies have been required to publish data open by default since a 2013 executive order, codified permanently by the OPEN Government Data Act in 2019.

The infrastructure got good — fast APIs, reliable catalogs, (sometimes) careful metadata.

That's not the problem anymore.

The bar the community set: FAIR

Written for machines as much as people — which is about to matter.

FAIR gets data to the doorstep. It still doesn't answer a question.

The answer is in here somewhere

Somewhere in a 12-million-row CSV is the number a reporter, a researcher, or a neighbor came for.

Getting there takes real work:

• knowing which question to ask
• understanding what the numbers mean
• making sense of the answer

That gap is where most people stop.

"Published" ≠ "answered"

A dataset is published A person got their answer

Everything after this slide is about closing the gap.

What DKAN is

• The Drupal distribution for open-data catalogs, in active use since 2013. Inspired by CKAN (another open-data catalog platform), built on Drupal.
• Maintained by CivicActions. Many public agencies run their data portals on it.
• DKAN 4 shipped on Drupal.org in March 2026 — composer require drupal/dkan, like any module.

×

Two halves: Metastore & Datastore

The AI tools discover through the metastore and answer through the datastore.

Anatomy of a dataset page

• Everything up top is the metastore: title, topics, description, publisher, distributions
• The table is the datastore: 2,969 rows — sortable, searchable, queryable
• "Also available via the API" — everything a human sees here, a tool can fetch

The DatastoreQuery language — not raw SQL

{
  "conditions": [
    { "property": "region", "value": "Southeast", "operator": "=" }
  ],
  "properties": [
    "park_name",
    { "expression": { "operator": "sum",
        "operands": ["sightings"] }, "alias": "total" }
  ],
  "groupings": [{ "property": "park_name" }],
  "sorts": [{ "property": "total", "order": "desc" }]
}

The three jobs an end-user actually does

These are exactly the three things that stop people.

…map directly onto AI strengths

The pairing isn't a gimmick — AI models are strong at exactly the interpretive layer people get stuck on. The deterministic catalog underneath keeps them honest.

Two ways to close the gap

Same goal, two audiences — though A quietly depends on B: dkan_ai_query builds on a submodule that ships inside dkan_mcp_server.

Drupal has names for these two ways

The bridge between them: dkan_query_tools — one protocol-neutral service layer that both workstreams consume.

Drupal AI Initiative, June 2026: one roadmap, two workstreams.

Meet the chat widget

"Ask in plain English — which parks had the most bear sightings? — and get an answer and a table, grounded in real queries. Ask for a chart and you'll get one, inline."

• A Drupal block; vanilla-JS library (no framework)
• Sanitized markdown · interactive tables + CSV · inline Vega-Lite charts
• A status pill shows the work ("Calling query_datastore…"), then the answer lands

What just happened

• A Drupal AI Agent (ai_agents, max 10 loops) with ~16 function-call tools.
• Mandated workflow: find resource → get schema (exact, case-sensitive columns) → sample / distinct values → query.
• It never guesses column names, and never does math in its head — every number routes through a tool.

The charts are model-authored

Chart-type selection is the model's job — "matching data shape to a chart," the AI-strengths mapping in practice.

From a chat box to an agent

Now hand the whole site to an agent.

• MCP — Model Context Protocol. A standard way for an AI client (Claude Code) to discover and call tools.
• Anthropic's open standard (Nov 2024); now under the Linux Foundation's Agentic AI Foundation.
• Tens of thousands of public MCP servers. This module makes DKAN one of them.

38 tools, in groups

Same server, two connections

Any MCP-aware client works out of the box — the gate, not the client, decides what's visible.

Where the AI actually lives

Keys live in the Drupal key module, never in code.

Built on the Drupal AI ecosystem

The DKAN modules are a domain layer on community foundations — no bespoke LLM plumbing, no bespoke protocol code.

Build once, expose twice

↘    ↙

Each surface is a ~10-line adapter over the same service methods — a future protocol is an adapter, not a rewrite.

The permission gate from the demo sits below both transports — stdio and HTTP behave identically.

Keeping token usage reasonable

The whole design assumes tokens cost money.

Cost & latency

Grounding, refusals, sanity flags

How we know it works

• The system prompt is a versioned file (v13 today) — not a config blob tweaked in prod
• 30 cases run against the live agent: expected answers, expected refusals, forbidden patterns
• Every stored answer records its prompt version + tool calls — regressions are traceable after the fact

Who can write, and what's switched on

Open data is in its own transition

Like the developer learning to work with AI agents, open data is mid-transition too.

The question isn't whether these tools replace the platforms we've built.

It's whether we can help the data we've already published finally reach the people it's for.

Three things to take home

1. The gap is interpretation, not infrastructure.
2. The catalog keeps the AI honest; the AI makes the catalog reachable.
3. Determinism where it counts, AI where it helps.

Questions?

Demo site is live — throw "can it do X?" questions at it.