io.github.IvandeMurard/paris-compass-mcp

The question

You are standing in front of an empty shopfront. The agent says the street is lively.

Compass answers on six axes at once.

Each of those arrives with its source, its date and its confidence level.

Note the wording, because the product keeps it too: air quality is measured, noise is modelled. Noise and footfall are proxies, labelled as such on screen and here.

Three that deserve more than a line

Has a kitchen ever been here? If a restaurant occupied the unit, the extraction, grease trap and power are probably already in. If not, creating them runs into tens of thousands and needs the building's agreement. The most expensive question this data answers — before you travel.

Is this a graveyard or a good street? A street that churns is not a street that dies: units can turn over and refill. And turnover only means something against its own trade, and against its own quartier — the same trade does not hold equally from one neighbourhood to the next. Compass measures both across three censuses.

Those figures are not printed here. They rest on the 2017 and 2020 vintages, whose licence has not been cleared, and the server withholds them from an anonymous caller. This page follows the same rule as the product: a rate derived from a withheld vintage is withheld too.

What did people actually pay? 25 661 goodwill sales published with their price since 2015. For the 5 971 tied to a single shopfront, the median Paris fonds changes hands between 160 000 and 170 000 € — a range, not a point, because goodwill prices are declared in round numbers: the median sits on a step and moves in jumps, so a single figure would be precise about something the source is not. The trade decides almost everything.

One address makes the point. At 10 rue Mandar, trace_premise returns six goodwill sales between 2017 and 2024, from 87 500 to 450 000 € — retrieved 7 October 2026. Each one is probable, not established: two shopfronts share that street number, and the notice does not say which was sold. The 2017 and 2020 census rows come back withheld, the 2023 one established.

How a number earns its place

Every figure arrives with its source, its licence, its date — and how sure it is.

flowchart LR
    A["APUR BDCom<br/>3 censuses"] --> M["Address<br/>matching"]
    B["BODACC<br/>sales + insolvencies"] --> M
    C["INSEE Sirene<br/>establishments"] --> M
    M --> V{"Does the source<br/>name THIS unit?"}
    V -->|"names it directly"| E["established"]
    V -->|"two sources agree,<br/>neither names it"| F["corroborated"]
    V -->|"address shared by<br/>several shopfronts"| G["probable"]
    V -->|"source is silent"| H["undetermined"]
    E --> O["Every figure carries<br/>source · licence · vintage · method"]
    F --> O
    G --> O
    H --> O
    O --> UI["Map and unit file"]
    O --> MCP["MCP server,<br/>for an agent"]

Four levels, never a percentage. A confidence score out of 100 would be exactly the kind of unverifiable number this product refuses. The level is computed from columns that already exist, and every row carries the reason that produced it.

Composition across the corpus, measured 6 October 2026 — this is the quality metric, and improving means moving these four numbers leftward:

That 37.0% is structural, not laziness: BODACC names an address, BDCom names a unit, and 69% of units share their street number. No public data will say which of eight shopfronts was sold.

A gate runs the whole corpus against 56 invariants, 24 frozen baselines and 8 hand-verified chronologies before anything ships, and again every day on a schedule — invariants counted 6 October 2026 with grep -c '^-- @invariant ' eval/invariants.sql. Most of them check what the functions return; one checks what they are — a function exposing an observed column must be SECURITY DEFINER, because row-level security silently turns a withheld row into "never surveyed".

The badge above counts invariants; it does not say the gate passes. The reds currently open are the issues labelled porte-rouge.

Two founding constraints

Most commercial-property tools describe the unit — floor area, rent, photos — and leave you to infer the rest. Compass sells neither coverage nor granularity: it sells interpretation, and context is the product.

An interpretation is worth only what you can check. The tools in this market promise answers grounded in audited data — that names who verified. Compass promises figures re-derivable from a cited public source — that names who can verify: you. An internal audit is a promise. A cited source is a test anyone can rerun, including against us.

If a number cannot be re-derived from a cited public source, it is not shown.

No landlord declarations, no scraped listings, no proprietary estimate, no score invented to fill a gap. Missing data is displayed as missing — n/a, never 0.

If two units on the same street get the same verdict, Compass has said nothing.

The useful granularity is the street segment, sometimes the side of the pavement. An indicator that does not vary at that scale describes general context; it does not settle a decision.

Who it is for

The entrepreneur — the shopkeeper, restaurateur, craftsperson or franchisee who decides where to open. Once or twice in a working life, on a lease that binds for years.

An agent — an LLM asking the same question through an MCP server. Same scoring core, same traceability requirement, different output: JSON and a chain of thought instead of a map.

Not brokers. A broker qualifies dozens of locations a month and needs portfolios, bulk comparison, national coverage. Every one of those makes the product heavier for someone studying a single address in depth. Serving both serves neither — a broker who wants Compass gets what the agent gets: the API.

What it refuses

Every refusal buys something back. The trade is the point.

Status

Honest labels, in the sense that built means the code runs and the gate passes — not that it is deployed.

Where every number comes from

Scope

Paris intra-muros first, Île-de-France next. Depth over breadth: the sources that carry the value are local.

Stack

React 18 · TypeScript · Vite 5 · Tailwind + shadcn/ui · Leaflet (direct, no wrapper) · TanStack Query · Supabase + PostGIS

src/core/ is pure — no fetch, no React, no DOM. That is what lets the same scoring serve a browser, a test runner and an MCP server.

git clone https://github.com/IvandeMurard/paris-compass.git
cd paris-compass
npm install
npm run dev

Licence and attribution

The code is Apache-2.0. The data is not.

That distinction matters more than it looks on a product built entirely from open data. The licence covers what is in this repository — the scoring core, the ingestion pipeline, the interface. It grants nothing over the datasets, which keep their own terms:

BDCom 2017 and 2020 are not redistributable at all. Their licence differs from the 2023 vintage, which is why the database carries a publicly_redistributable flag per vintage and withholds their content from an anonymous caller — content and absence alike, so that withholding leaks nothing either.

This repository contains no data extract, and must not. It contains the code that fetches, models and cites the data.

Made by Ivan de Murard.

Interested in the approach, or in the MCP layer? Open an issue.