Open-Source Tool

An open-source MCP lets your AI agent shop two Danish grocery stores, fill the cheapest cart, and never press pay — every regional retailer without an agent connector is next

An open-source MCP lets your AI agent search two Danish grocery stores, compare a whole basket with delivery fees, fill the cheapest cart — and you always press pay. Most regional retailers still have no agent front door.

Published Nov 6, 2026

Free check · public files only

Is your store or SaaS agent-ready?

Paste a public homepage. We read a few public files — robots.txt, llms.txt, markup, and whether a pricing page or agent endpoint is in the open — and show a short score right away. The full report unlocks when you leave an email. That signup is the same newsletter as everywhere else on this site: we email a confirmation link, and you confirm on our site.

This is a snapshot of public files, not a certification, a rank, or a prediction of sales. We do not log in, submit forms, call any tool on an MCP server, or store the page content. Results can be cached for 24 hours per domain.

The Wow

What someone built

handle-mcp (MIT) talks to nemlig.com and BilkaToGo. Tools cover product and recipe search, weekly offers, whole-basket compare, cart read/write, delivery slots and past orders. We checked the README, tools list and ADR on Nov 6, 2026.

Open-Source Tool

Original repository
handle-mcp on GitHub (MIT)
License
MIT (Copyright 2026 Niclas Johansen). Free to fork, modify and sell connectors built from it; keep the copyright notice
What you need
  • macOS, Linux or Windows x64; uv; Python 3.12+ (fetched by uv if needed)
  • An MCP client (Claude Code, Claude Desktop, Codex, Cursor, etc.)
  • A nemlig.com or BilkaToGo account with home delivery for cart writes; offers on nemlig work with no login
Install
uvx --from git+https://github.com/NJohansen/handle-mcp@v1.2.1 handle-mcp login nemlig && claude mcp add handle -- uvx --from git+https://github.com/NJohansen/handle-mcp@v1.2.1 handle-mcp

Why it's remarkable

Payment code does not exist. An architecture decision record says the human always pays, and a safety test fails if order or pay endpoints appear. Cart changes are reversible; orders are not — so the agent stops at the cart.

The Opportunity

The business it reveals

Every regional grocer, pharmacy or pet store without an agent connector is an opening. Fork the MIT pattern for one retailer in your country, or sell a household shopping agent that still checks out on the store's site.

Who would pay

Regional grocery chains and specialty retailers first; later, busy households on a monthly shopping-agent plan.

The Flip

3 ways you could turn this into a business

Pharmacy / pet / hardware cart MCP for local specialty retail

Who pays
Specialty retailers with online catalogs and repeat household buyers
How it would earn
Setup plus monthly for an agent that refills the usual cart and flags offers (our hypothesis)
Why it might work
Same compare-and-fill loop, narrower catalog, sticky repeat orders; handle-mcp's roadmap already points at more store types.
What already exists
Store apps' own reorder buttons; general shopping assistants that can't write to the cart.

Household meal-plan shopping agent (you still pay)

Who pays
Busy households that already buy online groceries
How it would earn
A monthly subscription for budget meal plans that fill the cheapest cart; the subscriber always pays the store (our hypothesis)
Why it might work
Taste and budget live in a skill on top of the server; compare_basket already returns cheapest-first with fees.
What already exists
Meal-kit brands; store apps' own meal plans; manual list sharing.

Why we'd try "Agent grocery connectors for one country's regional chains" first: Retailers already spend on apps and delivery; a demo that fills a real cart from Claude is a sales conversation, and the MIT + human-pays pattern is ready to fork.

Build It

How you could build it

An MCP server for one regional retailer that exposes search, offer list, basket compare and cart quantity tools only — no order or pay — plus a short skill that encodes the shopper's budget and dietary rules.

First step: Install handle-mcp from the v1.2.1 tag, ask your assistant to list this week's offers and compare a five-item basket across both stores, then open the cart yourself. Next, pick one grocer in your country and write down whether any public or partner API could support the same three tools: search, compare, set cart.

Agents and tools

  • handle-mcp as the MIT reference implementation
  • uv / Python
  • Claude Code, Codex or Cursor
  • The retailer's website or partner API (with permission)

Setup

  1. Run handle-mcp against nemlig/BilkaToGo and walk through compare_basket and set_cart_quantities.
  2. Read ADR 0001 and keep the same human-pays rule in your fork.
  3. Map one local retailer's catalog endpoints and terms of use.
  4. Implement search, compare and cart-write tools only; add a safety test that fails if order/pay appears.
  5. Demo to one retailer or five households with them paying on the store site.

What the agent does

  • Searches products and offers
  • Compares baskets and fills the cart
  • Drafts the connector code and skill prompts

What you do

  • Chooses the retailer and reads the terms
  • Sells the connector or household service
  • Reviews carts and supports breakages when APIs change

Your first customer: A friend who already shops online groceries, or one regional category manager shown a live compare of their basket vs a competitor.

Test demand first: Run the prompt on five grocers in one country this week and demo handle-mcp's compare_basket pattern on a screen recording to one retailer contact. Success: one asks what a pilot connector would cost.

Your starter instructions

Paste this into the AI agent you use (ChatGPT, Claude, Gemini, Grok Bot or another) as your first message. Copying it starts work in your own agent; it doesn't create or deploy a business by itself.

Get a launch plan

Running it

What stays human: Picking the retailer and country, reading the store's terms of use, testing login and cart writes safely, selling to the retailer or first households, and checking every cart before anyone pays.

Startup cost (estimate)
Under $25 to try (estimate) (Estimate: the MIT server runs locally via uvx; you need an AI plan you may already pay for and a free store account. Hosting a connector for a retailer later is a small hobby tier)
Running cost
Local runs are free beyond your AI plan; a hosted connector is single dollars a month plus your time (estimate) (Estimate: cart and search calls are light; hours unmeasured. The unofficial store APIs can break, which is the real keep-alive cost)
How much the agent runs
Agents run one core part; people run the rest.

Pricing your own version? Use the free startup-cost worksheet.

The evidence

What the source shows: GitHub repo NJohansen/handle-mcp (MIT, tagged v1.2.1), README tools table, ADR 0001, and site handlemcp.dk — checked Nov 6, 2026. No revenue is published.

What's ours: The connector studio, specialty-retail flips and household subscription are our applications of the pattern.

Worth planning around

  • Unofficial private website APIs can break any day; the README says so.
  • Cart writes need a real store login; Windows on ARM is not supported yet.
  • Nothing here shows retailers or households paying for this yet.

Sources

Sources reopened Nov 6, 2026. Spot a mistake? Tell us and we'll correct it in public.

Keep reading

The Daily Paycheck: one remarkable AI business at a time, and exactly how you could build around it. Free; unsubscribe anytime.

Get new AI business discoveries by email.