mcpField
A
Agentic House Search

Agentic House Search

Remote
agentic-house-search.vercel.appBusiness & Marketing1Updated 6d ago

About Agentic House Search

Agentic House Search is an MCP server for researching UK neighbourhoods through postcode reports and comparisons. It combines 45 government open-data sources and reports geographic coverage and comparability caveats alongside results.

Key features

  1. 01Postcode reports across 11 categories
  2. 02Geographic lookup for a postcode
  3. 03Compare two to five postcodes
  4. 04Search and inspect 45 datasets
  5. 05Coverage and comparison caveats

Use cases

  1. 01Research an area before moving
  2. 02Compare broadband and transport across postcodes
  3. 03Find datasets for property research

MCP support

The server exposes postcode research and dataset-registry tools to MCP clients over hosted Streamable HTTP or local stdio and HTTP. It can be connected through its published endpoint or run locally.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://agentic-house-search.vercel.app/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Packages
npmagentic-house-search0.1.3

Tools

5
ParameterTypeRequiredDescription
postcodestringYesA UK postcode, with or without a space, e.g. "SW11 1AA" or "m11ae"
categoriesarrayNoWhich sections to fetch. Omit for all eleven. Fetching only what you need is faster and much smaller — e.g. ["crime","schools"] for a family-safety question.
include_sourcesbooleanNoAppend the publisher and licence of each dataset used (default false)
response_formatenumNo'markdown' to read, 'json' for the full structured factsmarkdown | json

Details

From the project README

agentic-house-search

An MCP server for UK neighbourhood research. Give it a postcode, get back what the government's own open data says about that place — demographics, crime, deprivation, prices, fibre, 5G, noise, transport, amenities, schools and planning constraints — plus a searchable registry of the 45 datasets underneath.

It is the postcode report with an agent-shaped front door. Both run the same provider modules, so a threshold or a caveat is written once and shows up in both.

Connect

Nothing to install, no account, no API key. Paste this URL wherever your client asks for a connector, custom integration or MCP server URL:

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

The connect page has a copy button, one-click buttons for Cursor and VS Code, and the same instructions per client.

Claude Code:

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

Anything that connects by editing JSON — Claude Desktop (claude_desktop_config.json), a project .mcp.json, and most others:

{
  "mcpServers": {
    "agentic-house-search": {
      "type": "http",
      "url": "https://agentic-house-search.vercel.app/mcp"
    }
  }
}

GET /health says whether the endpoint is up, which is more useful than a client that only reports "connection failed".

Or find it by name

This server is listed in the official MCP registry, so a client with registry support can add it without being given a URL at all. Search for agentic-house-search, or use the canonical name:

io.github.asturksever/agentic-house-search

Anything resolving it programmatically can read the entry straight from the registry API — it carries both install paths, the hosted endpoint and the npm package, so a client can pick whichever it supports:

curl -s "https://registry.modelcontextprotocol.io/v0/servers?search=agentic-house-search"

The listing's version tracks releases, so it is worth checking it against /health if a client reports something unexpected — a stale listing means the registry entry was not republished after a release, not that the server is wrong.

Or run it yourself

Free, ungated and with no dependency on the hosted deployment. Requires Node 20+.

npx -y agentic-house-search              # stdio
npx -y agentic-house-search --http       # streamable HTTP on 127.0.0.1:8848
{
  "mcpServers": {
    "agentic-house-search": {
      "command": "npx",
      "args": ["-y", "agentic-house-search"]
    }
  }
}

Tools

ToolWhat it does
postcode_reportEleven categories for one postcode. Filter with categories to keep responses small.
postcode_lookupGeography only — coordinates, local authority, ward, constituency, LSOA/MSOA/OA codes, police force, deprivation rank. One fast call.
postcode_compareTwo to five postcodes side by side on chosen categories, with comparability caveats.
postcode_search_datasetsSearch the 45-dataset registry by text or category, paginated.
postcode_get_datasetOne registry entry in full: endpoint, API docs, licence, coverage, cadence.
Examples
"What's SW11 1AA like?"
  → postcode_report(postcode="SW11 1AA")

"Which of these three has the best broadband and transport?"
  → postcode_compare(postcodes=[...], categories=["broadband","transport"])

"Where would I get EPC data for a property?"
  → postcode_search_datasets(query="EPC") → postcode_get_dataset(id="epc")

What the numbers mean

The point of this server is that every figure states what it actually describes. Agents summarising it should carry that through:

  • Geography varies by source. A census figure describes an LSOA — a neighbourhood of roughly 1,500 people, not an address. A crime count describes a 1 km square. Ofcom mobile coverage describes an entire local authority, because that is the finest grain Ofcom publishes; two postcodes in the same authority will always show identical mobile figures. Ofcom broadband is per postcode.
  • Coverage varies by UK nation, and is stated rather than hidden. Census tables are England & Wales; data.police.uk excludes Scotland; the Planning Data platform, Defra noise and the DfE school register are England-only. Those categories come back as out_of_coverage with the reason and a link to the devolved equivalent — never as a zero or an empty result.
  • Deprivation ranks are not comparable across nations. England, Wales, Scotland and Northern Ireland each rank their own areas against their own index over a different number of areas. The index and its size are always returned; postcode_compare refuses to let a cross-nation comparison pass without a caveat.
  • Police data depends on each force submitting. A very low count in a built-up area is more likely a gap than a quiet street, and the report says so when the count is implausibly low.
  • unavailable means not built yet, not "none" — currently the Defra noise extract, which needs a polygon join that has not been run.

This is not a survey, a valuation or a conveyancing search.

Configuration

VariableDefaultPurpose
AHS_BASE_URLthe published siteWhere to read the registry and pack extracts. Point it at http://localhost:8000/ to develop against a local checkout.
AHS_JS_ROOTunsetWhere the shared js/ provider modules live. A hint, not a requirement: with it unset the server checks next to the compiled output (the published package) and then ./js (a bundled function), and fails with the list of directories it tried. Only set it when neither is right.
ALLOWED_ORIGINSnoneComma-separated Origin allowlist for HTTP mode. Requests carrying any other Origin are rejected with 403.
API_KEYSunsetHTTP mode only. Comma-separated key or key:pro. Unset means every caller is anonymous and nothing is rejected.
RATE_LIMIT_ANONYMOUS60/hourHTTP mode only. A courtesy limit so one runaway agent cannot burn the upstream fair-use budgets.
RATE_LIMIT_PRO1000/hourHTTP mode only.

HTTP mode binds to 127.0.0.1 by default and is stateless: a fresh server per request, so it scales horizontally with no session affinity.

Running it yourself is free and ungated, and stays that way. stdio has no limits at all, and --http with no API_KEYS set is open. The rate limiting exists so that a shared deployment is a good neighbour to the government APIs underneath, not to nudge you toward a paid tier. There isn't one. See COMMERCIAL.md for where that boundary sits and what would have to be true before any of it were sold.

The hosted endpoint

https://agentic-house-search.vercel.app/mcp is this same package, built from this repository, deployed as a serverless function (api/mcp.mjs, vercel.json) on the same host as the website, which scripts/stage-site.mjs assembles into public/ at build time. It exists so that connecting takes a URL rather than a config file. It is unauthenticated because there is nothing to authenticate: every source is public open data and the server holds no per-user state.

Both HTTP hosts share src/http.ts, so the hosted endpoint and your own --http cannot drift apart in how they speak the protocol.

Deploying it

This repository is the deployment. Vercel builds from GitHub: connect the repo once, and every push to main redeploys the endpoint. There is no CLI step, no separate copy of the source and nothing to remember to run — the same push that updates the website updates the MCP server, and a revert reverts both.

Setup is once, in Vercel's Add New → Project → Import Git Repository. Two things matter:

  • Name the project agentic-house-search. vercel.json carries every other setting, but not the project name, and the project name is what the URL is made of. Any other name and the documented URL is a lie — change it in connect.html, server.json, both READMEs and index.html, or alias a domain onto it.
  • Leave the framework preset on "Other." The build command and output are already in vercel.json; a preset would override them.

The build runs npm ci && npm run build in mcp/, and AHS_JS_ROOT=js tells the bundled function where the provider modules landed. .vercelignore keeps the website's own files out of the upload, since Pages serves those and / here redirects there.

npm run smoke:http drives the same entrypoint locally and runs in CI on every push, so a deployment that would break should go red in Actions first.

server.json is the entry for the official MCP registry, which is how a client can offer this server by name rather than by URL. Publish it with the registry's own CLI (brew install mcp-publisher), from the repository root:

mcp-publisher login github
mcp-publisher publish

Publish to npm first. The registry proves you own the package you point it at by reading mcpName out of the published npm package and comparing it with server.json#name — so it validates what npm already has, not what is in this repo. A version listed in server.json that npm has never seen fails, and because npm will not accept the same version twice, the fix is another bump. npm run check:versions compares both fields locally and runs in CI.

Development

npm install
npm run build      # tsc, then copy ../js into dist/js
npm run smoke      # drives the built server over stdio and checks every tool
npm run smoke:http # drives the serverless entrypoint the hosted endpoint uses
npm run inspect    # MCP Inspector

npm run build copies the repo's js/ provider modules into dist/js. That copy is deliberate: the alternative is reimplementing eleven providers, their thresholds and their coverage gates in TypeScript, which is exactly how the website and the server would start disagreeing about the same postcode.

evaluation.xml holds ten verified questions for testing whether a model can actually use these tools. Every answer comes from fixed geography, a dated statistical release or the registry — never a live figure that moves monthly.

Licence and attribution

This server is CC0. The data is not: it is public sector information under the Open Government Licence v3.0, plus OS and Royal Mail rights in the postcode geography, and OpenStreetMap contributors (ODbL) for amenities. Every response carries the attribution — please keep it attached.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.