agentic-house-search-data-registry

# Agentic House Search

**Enter a UK postcode, get what government open data actually says about that
neighbourhood.** Demographics, crime, deprivation, prices, fibre, 5G, noise,
transport, amenities, schools and planning constraints, assembled live from the
publishers' own APIs.

Three ways in, all free and open:

| | |
| --- | --- |
| **[Postcode report](https://agentic-house-search.vercel.app/report.html)** | The website. Type a postcode, read the answer. |
| **[MCP server](mcp/README.md)** | Paste one URL into Claude, ChatGPT or any MCP client and your agent can do the same. [Connect an agent →](https://agentic-house-search.vercel.app/connect.html) |
| **[Dataset registry](https://agentic-house-search.vercel.app/registry.html)** | The 45 open datasets underneath, with endpoints, licences and coverage. |

## Why this exists

UK property intelligence sites (CrystalRoof, PropertyData and others) are
largely assembled from public data. This project maps out that raw material, the
Open Government Licence datasets and free APIs underneath, and then shows what
you can build with it: a working report, and an agent-ready interface to the
same thing.

Two rules it holds itself to, which is most of the point:

1. **It never invents a comparison.** Every band word ("above the UK average")
   comes from a threshold table and always sits next to the figure it describes.
2. **It never pretends to cover ground it does not.** A Scottish postcode is
   told that `data.police.uk` excludes Scotland, with a link to Police Scotland,
   rather than being shown a zero. Deprivation is computed against each nation's
   own index and denominator.

## Contents

| File | Description |
| --- | --- |
| [`data/registry.json`](data/registry.json) | Canonical machine-readable registry |
| [`data/registry.csv`](data/registry.csv) | Flat table for spreadsheets (generated) |
| [`index.html`](index.html) | Landing page — what this is, the figures, the category index |
| [`registry.html`](registry.html) | Searchable catalogue of every entry |
| [`report.html`](report.html) | Postcode report — pulls live values out of these sources for one postcode |
| [`js/`](js) | The provider layer: one module per category, plus the thresholds and coverage gates. Shared by the website and the MCP server. |
| [`mcp/`](mcp/README.md) | MCP server, published as [`agentic-house-search`](https://www.npmjs.com/package/agentic-house-search) |
| [`packs/`](tools/packbuild/README.md) | Pre-built extracts for the sources that publish bulk files instead of an API |
| [`tools/`](tools) | `gen_derived.py` (regenerates the CSV and the README tables), `validate_registry.py`, `check_provider_ids.py`, `packbuild/` |

Both pages fetch `data/registry.json` at runtime, so they need to be served over
HTTP rather than opened from disk: `python3 -m http.server 8000`, then
<http://localhost:8000/>.

## Postcode report

[`report.html`](report.html) is the registry pointed at a single postcode: enter
one and it fetches the real values out of these sources and writes a plain
paragraph per category — who lives here, crime, deprivation, prices, fibre, 5G,
noise, transport, amenities, schools, planning constraints.

It runs entirely in the browser. Most sources are called directly (postcodes.io,
data.police.uk, Nomis, HM Land Registry, Planning Data, the Environment Agency,
TfL, Overpass); the three that publish bulk files instead of an API — Ofcom
fibre, Ofcom mobile, the DfE school register — are pre-joined into `packs/` by
[`tools/packbuild`](tools/packbuild/README.md).

Results open with a tile per category — the headline figure and how it rates —
then one card each with the numbers behind it. Ratings are colour-coded, but
never by colour alone: each carries a glyph (▲ good, ● typical, ◆ worth a look,
▼ needs attention) and the band in words, so the meaning survives greyscale,
printing and colour blindness. Categories that cannot be judged good or bad — a
house price is neither — stay deliberately uncoloured.

Two things it will not do. It never invents a comparison: every band word
("above the UK average") comes from `js/thresholds.js` and always sits next to
the figure it describes, which is in the table below the paragraph. And it never
pretends to cover ground it does not — a Scottish postcode gets an explicit
"data.police.uk does not publish Scottish data" with a link to Police Scotland,
not a blank card. Deprivation deciles are computed against each nation's own
index and denominator.

An **AI summary** is optional and off by default. There is no server here to
keep an API key in, so it uses one you paste yourself, stored in your own
browser and sent to nothing but `api.anthropic.com`. The model is given only
the computed facts, and every number it writes back is checked against them —
anything untraceable is flagged rather than shown quietly. Without a key there
is a "copy the prompt" button, and the report is complete either way.

Local development (both pages fetch `data/registry.json`, which `file://`
blocks):

```bash
python3 -m http.server 8000   # then http://localhost:8000/report.html?postcode=SW111AA
```

## MCP server

[`mcp/`](mcp/README.md) exposes all of this to AI agents over the Model Context
Protocol — five tools: a full postcode report, a fast geography lookup, a
side-by-side comparison of up to five postcodes, and search/fetch over the
registry itself.

There is nothing to install. Paste this into your client's connector or custom
integration field:

```
https://agentic-house-search.vercel.app/mcp
```

Or add it in one command:

```bash
claude mcp add --transport http agentic-house-search https://agentic-house-search.vercel.app/mcp
```

Or, in a client that supports the
[official MCP registry](https://registry.modelcontextprotocol.io), add it by
name without a URL at all — search `agentic-house-search`, or use
`io.github.asturksever/agentic-house-search`.

The [connect page](https://agentic-house-search.vercel.app/connect.html)
has the same thing with a copy button and per-client instructions. The hosted
endpoint is free, needs no account or key, and carries a courtesy rate limit
that exists to protect the government APIs underneath — not to sell you
anything. Running it yourself has no limit:

```bash
npx -y agentic-house-search          # stdio, published on npm
npx -y agentic-house-search --http   # your own streamable HTTP endpoint
```

It runs the **same** provider modules as `report.html` rather than a
reimplementation, so a threshold, a coverage gate or a caveat is written once
and both surfaces agree. See [`mcp/README.md`](mcp/README.md).

## Schema

```json
{
  "id": "police-street-crime",
  "dataset": "Street-level crime, outcomes and stop & search",
  "publisher": "Home Office / UK police forces (data.police.uk)",
  "link": "https://data.police.uk/",
  "api": "https://data.police.uk/api/crimes-street/all-crime?lat={lat}&lng={lng}&date={YYYY-MM}",
  "api_docs": "https://data.police.uk/docs/",
  "format": "JSON REST API (no key required); CSV bulk downloads",
  "licence": "Open Government Licence v3",
  "coverage": "England, Wales & Northern Ireland",
  "update_frequency": "Monthly",
  "questions": [
    "How safe is this postcode / street compared with the borough or national average?",
    "What are the most common offence types here (burglary, ASB, drugs, vehicle crime)?",
    "Is crime trending up or down over the last 12-36 months?",
    "Where are the local crime hotspots and what happened to each case (outcomes)?"
  ],
  "category": "Crime"
}
```

## Datasets by category

### Crime

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Street-level crime, outcomes and stop & search](https://data.police.uk/) | Home Office / UK police forces (data.police.uk) | `https://data.police.uk/api/crimes-street/all-crime?lat={lat}&lng={lng}&date={YYYY-MM}` | Open Government Licence v3 |

### Demographics

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Census 2021 (ethnicity, religion, age, household composition, tenure, occupation, qualifications)](https://www.ons.gov.uk/census) | Office for National Statistics | `https://api.beta.ons.gov.uk/v1/population-types` | Open Government Licence v3 |

### Demographics / Overview

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Output Area Classification (OAC 2021) & London Output Area Classification (LOAC)](https://data.london.gov.uk/dataset/london-area-classification) | ONS / Greater London Authority (London Datastore) | `https://data.london.gov.uk/api/action/package_show?id=london-area-classification` | Open Government Licence v3 |

### Affluence

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Small area model-based income estimates (MSOA) + approximated social grade & qualifications (Census tables TS062/TS067)](https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/earningsandworkinghours/datasets/smallareaincomeestimatesformiddlelayersuperoutputareasenglandandwales) | Office for National Statistics | `https://www.nomisweb.co.uk/api/v01/dataset/def.sdmx.json` | Open Government Licence v3 |

### Affluence (Deprivation)

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [English Indices of Multiple Deprivation (IMD 2019)](https://www.gov.uk/government/statistics/english-indices-of-deprivation-2019) | MHCLG (Ministry of Housing, Communities & Local Government) | `https://opendatacommunities.org/sparql` | Open Government Licence v3 |

### Property prices

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Price Paid Data (every residential sale) + UK House Price Index](https://www.gov.uk/government/collections/price-paid-data) | HM Land Registry | `https://landregistry.data.gov.uk/landregistry/query` | Open Government Licence v3 (address fields have extra conditions) |

### Property prices (Rents)

| Dataset | Publisher | API | Licence |
| --- | --- | --- | --- |
| [Price Index of Private Rents