ShelfSignal

The open store index for AI agents.

Local merchants list what they actually sell. AI agents find real food and real products nearby, score how likely each item is actually in stock, and check a validation receipt before they send a human across town. Merchants pay one flat price to get listed. Nobody takes 30% of their food.

DoorDash and the delivery apps built a closed index: they own the customer, scrape or rent the menu, charge the merchant 15-30% per order, and still cannot tell you whether the kitchen has the thing you want. ShelfSignal is the opposite shape:

• Merchant-owned listing. The store submits its own stuff, its own prices, its own stock. The listing belongs to the merchant, not the platform.
• Built for agents, not ads. One MCP server any agent can query: find stores, find items, check stock, validate the listing. No sponsored placement, ever.
• Stock is a scored signal, not a claim. Every item carries a last-confirmed date. The likelihood engine turns that evidence into a 0-100 score with printed reasons, so agents stop bluffing "they have it" and start saying "confirmed today, likely in stock."
• Validation receipts. Every listing carries checks an agent can verify: complete profile, real prices, fresh inventory, declared fulfillment. Tier: VERIFIED, LISTED, STALE, or INCOMPLETE.
• Flat money. $100 Starter Store, one time, listed and validated. No per-order commission. See docs/PRICING.md. Merchants can join by email: docs/EMAIL-INTAKE.md.

Quick start

npm install
npm run build
npm test

node dist/cli.js coverage
node dist/cli.js search "birria" --city "San Francisco"
node dist/cli.js stock sf-mission-taqueria-demo "Birria tacos"
node dist/cli.js nearby 37.7599 -122.4148 --radius 5
node dist/cli.js ping sf-sunset-bakery-demo "Cardamom knot" in
node dist/cli.js validate sf-mission-taqueria-demo
node dist/cli.js doctor
node dist/cli.js mcp        # run the MCP server (stdio)
node dist/cli.js serve      # run the hosted registry API (http://0.0.0.0:8787)

Hosted registry API

A read-only HTTP API over the same registry, for agents and clients that do not speak MCP:

curl localhost:8787/v1/stores?city=San%20Francisco
curl localhost:8787/v1/stores/sf-mission-taqueria-demo
curl "localhost:8787/v1/search?q=birria"

Only listings in the live state are served: listings still in the intake queue are excluded from list and search, and their detail route returns 404. Full reference: docs/API.md.

MCP tools

| Tool | What it does | |---|---| | search_stores | Find stores by text, city, category, or type (with open-now signal) | | find_nearby | Stores near a lat/lng point, nearest first, with distance and open-now | | search_items | Find a food or product across stores, every hit carrying its stock verdict | | get_store | Full listing: hours, address, fulfillment, items, stock | | check_stock | Stock-likelihood verdict (score, band, reasons) for one item | | validate_store | Validation receipt and tier for one listing | | submit_store | Validate a merchant intake and preview its listing + receipt | | list_coverage | Cities, categories, and store count in the index |

For merchants

You run a store in the city. For $100 you get listed where the agents look:

1. Fill the intake (10 minutes): who you are, where you are, your hours, and the stuff you sell with real prices. Template: examples/submission.example.json, field guide: docs/INTAKE.md.
2. We validate it and your listing goes live with its first receipt.
3. Update stock whenever it changes with one command or one message. Fresh listings rank as VERIFIED; stale ones decay honestly instead of lying to customers.

No commission. No tablet on your counter. No competing with your own listing.

How it works

• The registry is the source of truth: registry/stores.json (v0.1, git-backed, every change reviewable).
• The likelihood engine is deterministic: merchant-reported status, confirmation freshness, staple vs specialty. No model guesses.
• The management layer is freshness: listings decay, validation re-runs on every read, and merchants see exactly what an agent sees. Roadmap in docs/MANAGEMENT.md.
• Full design: docs/SPEC.md. Why this beats the delivery model: docs/VS-DOORDASH.md.

Bundled stores are sample data and are flagged demo: true in every response. Real merchants are never marked demo, and demo stores are never presented as real.

License

MIT