uk land registry mcp

three more tools do the plumbing: postcode → coordinates, district and ward; a status call that reports what is cached and what to fetch next; and the downloader for the bulk datasets. three bundled prompts wrap the common jobs — a due-diligence report on an address, an ownership investigation, a site analysis.

52 COATES AVENUE, TS4 3AQ — 2 registered sales
29 jul 2022   £186,100    —       —
20 apr 2026   £195,000   +4.8%   +1.3%/yr

→ hmlr_data_status   ocod cached · july 2026
→ hmlr_search_ownership_by_area
   title · proprietor · country of incorporation

• +210.6%over 22.87 years
• +5.1%a year, compounded
• −32%below the local index

the gap is the finding, not the explanation. a shortening lease, the condition of one flat, or simply a 1998 sale struck above the going rate would each produce this shape — and indexing from a single transaction inherits whatever that transaction got wrong. the server reports the divergence and stops there.

these 13 sales cannot be attributed to a single dwelling, because the sales are of flats and no saon was given, so this address may cover every flat in the building rather than one of them.

a tool that is wrong out loud is worse than one that is silent, because a table of percentages reads as fact. most of the work here was not adding tools — it was finding the places where confident output was fiction.

so the query builder emits the selective triples first, drops the type triple entirely — it alone matches every row — and keeps the optional projections last. searches too broad to be made selective are refused locally rather than sent off to time out. the ordering is asserted in the test suite, because nothing else would catch it going wrong.

1. 01

   check you have node 22

   the server is a node process your client starts on demand. nothing else to install — no docker, no python, no account.

   node --version   # v22 or later

   older node will build and then fail at runtime; the test suite uses native typescript stripping.

2. 02

   clone and build it

   one command, run wherever you keep repos. it compiles to dist/index.js — that file is what your client will run.

   git clone https://github.com/light-vp/uk-land-registry-mcp.git
   cd uk-land-registry-mcp
   npm install
   npm run build

3. 03

   point your client at it

   stay in the repo directory for the claude code line — $PWD fills in the absolute path for you. for claude desktop, paste the block into claude_desktop_config.json and swap in the real path, then restart the app.

   claude code

   claude mcp add land-registry -- node "$PWD/dist/index.js"

   claude desktop · claude_desktop_config.json

   {
     "mcpServers": {
       "land-registry": {
         "command": "node",
         "args": ["/absolute/path/to/uk-land-registry-mcp/dist/index.js"]
       }
     }
   }

   that is it for sold prices, the index and postcode lookups. ask “how has the middlesbrough housing market moved over the last five years?” and you should get a table back.

4. 04

   optional — turn on ownership and boundaries

   register free at use-land-property-data.service.gov.uk, then accept the licence for each dataset you want — access is granted per dataset, and a key that works for ccod will not fetch ocod until you accept ocod’s licence too. that one thing causes nearly every 403. add the key to the same config and restart.

   register for a key ↗

   claude code

   claude mcp add land-registry --env HMLR_API_KEY=your-key-here -- node "$PWD/dist/index.js"

   claude desktop · alongside "args"

   "env": { "HMLR_API_KEY": "your-key-here" }

   then ask it to download the overseas ownership dataset — smallest of the bulk files, and a good first one. the downloader always reports the file name and size before fetching anything, so you see the cost before you commit; ccod is several hundred mb. everything lands in ~/.hmlr-mcp.

the readme covers the rest — every tool, the failure modes worth knowing, and why a boundary polygon can never be joined to its owner using open data alone.