pier39-merchant-server

Drop-in negotiate.v1 backend for any storefront. You bring a catalog.json. The library brings the chat agent, the discovery files, and the rate limiting. Two minutes from pip install to a running negotiable backend, via the pier39-merchant CLI wizard.

Quickstart (recommended)

pip install pier39-merchant-server
pier39-merchant init

The wizard asks for your store name, your Anthropic API key, and where to load products from (sample data, generic CSV, or a Shopify export). It writes catalog.json, serve.py, and .env into the current folder. Then:

pier39-merchant serve     # run locally on port 8000
pier39-merchant deploy    # deploy to Fly.io (requires flyctl)

The merchant skill is bundled inside the package — no git clone required.

Manual / library use

If you'd rather wire it up yourself:

pip install pier39-merchant-server

Six lines, including the import:

from pier39_merchant_server import serve

serve(catalog_path="catalog.json", host="0.0.0.0", port=8000)

(The library auto-discovers the skill at ~/.claude/skills/pier39-merchant/ if installed there, otherwise falls back to the bundled copy. Override with PIER39_MERCHANT_SKILL_PATH env var or pass skill_path= directly.)

That gives you, on http://0.0.0.0:8000:

All public, all CORS-permissive (Access-Control-Allow-Origin: *).

Catalog format

A catalog.json looks like this — see examples/tiny_store/catalog.json for a working minimal example:

{
  "store": {
    "name": "Your Store",
    "city": "City, ST",
    "rep_name": "Casey",
    "tagline": "...",
    "policy_default": "Free shipping. 30-day return. ..."
  },
  "products": [
    {
      "id": "widget-pro",
      "kind": "widget",
      "name": "Widget Pro",
      "subtitle": "Short product line",
      "description": "Long description (model, accessories, etc.)",
      "year_made": 2024,
      "condition": "Brand new",
      "list_price": 199,

      "floor": 165,
      "levers": [
        "Free expedited shipping (~$15 retail value)",
        "Bonus accessory kit ($29 retail)",
        "Extended warranty (12 → 24 months)"
      ],
      "inventory_note": "..."
    }
  ]
}

The floor, levers, and inventory_note fields are hidden — they appear only in the merchant agent's private system prompt, never in any GET response. The library's /api/store/catalog endpoint strips them automatically.

Deeper control

If you need to customize, instantiate MerchantApp directly:

from pier39_merchant_server import MerchantApp

app = MerchantApp(
    catalog_path="catalog.json",
    skill_path="/custom/path/to/pier39-merchant",
    model="claude-opus-4-6",                    # default: claude-sonnet-4-6
    max_chat_starts_per_hour_per_ip=20,         # default: 8
    max_messages_per_chat=50,                   # default: 30
    max_message_length=3000,                    # default: 2000
    session_idle_ttl_seconds=7200,              # default: 3600 (1h)
)
app.run(host="0.0.0.0", port=8000)

You can also pull app.make_handler_class() and mount it under your own ThreadingHTTPServer / nginx / wsgi setup if you have an existing stack.

Required env vars

• ANTHROPIC_API_KEY — for the merchant chat agent
• PIER39_MERCHANT_SKILL_PATH (optional) — override skill auto-discovery

What you get for free

Any AI shopper agent that speaks negotiate.v1 can now negotiate at your store with no further coordination. The Pier39 MCP connector is one such shopper — install it in Claude Desktop and you can immediately negotiate at any negotiate.v1 storefront.

See also

• PROTOCOL.md — the formal protocol spec
• pier39-skills — the merchant + shopper Claude Agent Skills
• Atlas demo — full reference implementation

License

MIT.