Small Business Intelligence

A free, open source MCP server that teaches any AI how to research a small business — and, more usefully, where the public records actually are.

Nine of its tools ship methodology, not data. Each returns a rigorously structured framework — a research procedure, an output schema, a quality rubric, the traps — and the calling model executes the research itself, with its own tools and its own keys. Those nine call nothing at all.

Two ship the records. twin_cities_datasets and twin_cities_records answer from joined public records for the seven-county Minneapolis-St. Paul metro — parcels and lot lines, recorded sale prices, owners, rental licences, contamination files, business counts by trade, census tracts. Ask about one address and they answer about that address. They call brickandmortar.dev/api/export, which we run; no tool here calls any third party. Every answer is a true row count, at most six example rows, and a link to the complete file — never a file inline. See /privacy for exactly what those two transmit and what is kept.

That split is deliberate and dated: the server was built on "no remote calls to our servers" (2026-08-16), which rested on a thesis retired the next day — the join is the moat, and the data ships. Confirmed 2026-08-20; the nine are untouched.

Live endpoint: https://sbi-mcp.small-business-intelligence-mcp.workers.dev/mcp No auth, no account, no key. Add it as a custom connector and ask.

Why this exists

A capable model already knows how to reason about a small business. What it does not know is the operational trivia that lives in nobody's training data:

• that a county's parcel geometry arrives in survey feet in Kansas and metres in Minnesota, so a hard-coded threshold silently triples;

• that Esri's Touches predicate returns zero touching parcels rather than an error, so adjacency quietly becomes "this parcel touches nothing";

• that Kansas never records a sale price at all, so an hour spent looking for one is an hour spent looking for something that does not exist;

• that Google's Places API returns at most five reviews, relevance-ranked, so a sentiment trend computed from them is a real-looking number from a sample somebody else chose;

• that County Business Patterns suppresses small cells, so reading one as zero turns a thin market into an empty one.

Every one of those produces a plausible wrong answer rather than a failure. That is the whole problem with public data, and it is what this server is for.

The frameworks are the reasoning. src/tools/sources.ts is the map — where each record lives, how to reach it, and the specific way it lies. Every access pattern and trap in it was measured live against the agency's own endpoint while building a real two-metro property and review corpus, not recalled from training data.

Related MCP server: Scout Intel MCP

The tools

Start with data_source_atlas. It is the one that changes what the rest are worth.

What it will not do

Stated plainly, because the boundary is the design:

• No remote calls to us. Every endpoint the tools name is reached by your model, directly, with your own keys. Nothing routes through this server.

• No data. There is no corpus here, nothing cached, nothing to go stale.

• No account, no telemetry about you. The only thing recorded is an aggregate per-tool call counter with no identity attached. See PRIVACY.md.

• No claim to do financial diligence. Revenue, margin, private lease terms and the terms of a private sale are not public anywhere in the United States. The tools say so rather than substituting a proxy.

Connect it

Add the endpoint as a custom connector in Claude, or any MCP-compatible client:

https://sbi-mcp.small-business-intelligence-mcp.workers.dev/mcp

Then ask something real:

"Where would I actually find what 1420 Grand Ave in Saint Paul last sold for?"

"I want to know if Wichita has room for another dog daycare — what should I pull?"

"Run a business_teardown on Mucci's Italian in Saint Paul, MN."

Run your own

npm install
npm run typecheck    # tsc --noEmit
npm run dev          # wrangler dev — serves http://localhost:8787

wrangler dev runs against Miniflare's local KV simulation, so the usage ledger works out of the box with no Cloudflare account needed.

Verify it

# list every tool
npx @modelcontextprotocol/inspector --cli --server-url http://localhost:8787/mcp \
  --method tools/list

# call one
npx @modelcontextprotocol/inspector --cli --server-url http://localhost:8787/mcp \
  --method tools/call --tool-name data_source_atlas \
  --tool-arg question="what did this building last sell for" \
  --tool-arg place="Wichita, KS"

Or drop --cli for the interactive web UI (npm run inspector). All tools should list with readOnlyHint: true and an outputSchema, and a tools/call against each should return structuredContent matching it, with no isError.

Deploy

wrangler login                              # or set CLOUDFLARE_API_TOKEN
wrangler kv namespace create USAGE_LEDGER   # once — copy the id into wrangler.jsonc
npm run deploy

Ships to the free *.workers.dev subdomain — no DNS work, no paid plan.

A note on context cost

tools/list is ~43 KB (~10.8K tokens) and sits in the client's context for the whole session whether or not a tool is ever called. Most of that is the shared outputSchema serialised once per tool. If you fork this and add tools, read the comment at the top of src/tools/types.ts first — every .describe() in that schema is paid for once per tool, forever.

Repo structure

src/
├── index.ts        # Worker entry — routes /, /docs, /privacy, /mcp
├── server.ts       # createServer(): builds McpServer; TOOL_NAMES is the one list
├── env.ts          # Env (Worker bindings/vars) type
├── middleware/     # identity resolution, KV usage ledger, policy, withPolicy seam
├── oauth/          # OAuth 2.1 discovery handlers — written, not mounted
├── tools/
│   ├── sources.ts          # WHERE THE RECORDS ARE — parcel GIS, Census, state/local, reviews
│   ├── federal_sources.ts  # BLS series construction + six measured traps
│   └── *.ts                # one file per tool, sharing FrameworkPayload from types.ts
└── pages/          # landing (/), docs (/docs), privacy (/privacy)

Architecture and the SDK/transport decision: ARCHITECTURE.md.

Contributing

The most valuable contribution is a measured trap. If you pull a public source and it lies to you in a way that returns a plausible number instead of an error, that belongs in sources.ts — with how you measured it and when. Access patterns rot as agencies reorganise; a correction with a date on it is worth more than a new framework.

Please don't add anything that makes the server call an external service. The no-outbound-calls property is what makes it free to run and safe to trust.

Who made this

Built by Brick & Mortar — a small team in Saint Paul that maintains real local-market corpora (county parcel records, recorded sales, review panels, federal series) for its own products. The frameworks here are what we learned building those, including the traps that silently return a plausible wrong number.

This is a gift, not a funnel. It is not a demo of a paid product, there is nothing gated, and nothing here reports back to us.

License

MIT — see LICENSE.