Skip to main content
Glama

CartClaw

Let your AI agent use your own Amazon account: read orders and return deadlines, look up products, fill the cart. It can only buy after you click Place order on an approval page. cartclaw.dev

Not affiliated with or endorsed by Amazon. Amazon is a trademark of Amazon.com, Inc.

Everything runs on your machine. The agent drives a separate browser window with its own profile, signed in to your account once. No Amazon password or session leaves your computer, and your everyday browser is never touched.

What the agent can do

Tool

Does

orders

Order history (last30, months-3, year-YYYY) with each item's return window

returnable_items

Items you can still return, soonest deadline first

product

Title, price and availability for an ASIN

cart

What is in the cart

cart_add / cart_remove

Change the cart

checkout_request

Ask you to approve buying the cart

checkout_status

Whether you approved and the order went through

There is no tool that places an order.

Related MCP server: Amazon Cart MCP Server

How buying works

  1. The agent fills the cart and calls checkout_request.

  2. The server opens Amazon's checkout review page in its window and reads the total, items, address, card and delivery date. Nothing is placed.

  3. An approval page opens in your normal browser with those details, a screenshot of Amazon's page, and two buttons: Place order and Don't buy. If the card is wrong, pick another saved card and click Use this card: the server switches it on Amazon's checkout and the page shows the new total.

  4. If you click Place order, the server reloads checkout, switches back to the card you approved (Amazon puts its default card back whenever checkout reopens), checks that the total, items, address and card still match what you saw, and only then clicks Amazon's "Place your order". If anything changed, nothing is ordered.

  5. The request expires after 15 minutes. A newer request replaces an older one.

Setup

Needs uv, git, and Brave, Chrome, Edge or Chromium.

  1. Sign in to Amazon once. A window opens; tick Keep me signed in: uvx --from git+https://github.com/ucsandman/cartclaw cartclaw login

  2. Add it to Claude Code: claude mcp add --scope user cartclaw -- uvx --from git+https://github.com/ucsandman/cartclaw cartclaw

Other MCP clients can run the same uvx command as a local stdio server.

The browser window starts by itself when a tool needs it and stays open between sessions. You can minimize it.

Optional settings (environment variables): CARTCLAW_BROWSER (browser path), CARTCLAW_PROFILE (profile folder, default ~/.cartclaw/profile), CARTCLAW_CDP_PORT (default 9333).

Development

git clone https://github.com/ucsandman/cartclaw && cd cartclaw && uv sync. The marketing site at cartclaw.dev is static HTML plus one Vercel function in site/.

Site operations (2026-10-02):

  • A push to main deploys the site. The Vercel project's root directory is site.

  • Google Search Console and Bing verify through the google-site-verification and msvalidate.01 meta tags in site/index.html. Removing either tag revokes that verification.

  • Vercel Web Analytics loads from /_vercel/insights/script.js on every page.

  • Pro waitlist signups are stored as private files in the Vercel Blob store cartclaw-waitlist.

  • uv run pytest: parsers against saved Amazon pages (personal details scrubbed), the approval page, and the tool list. No network.

  • uv run python scripts/live_smoke.py: read-only check against your real account over the MCP protocol.

Limits

  • amazon.com (US) only.

  • Returns: return deadlines are read, but starting a return is not built yet.

  • The approval page lists saved cards only. To pay with a bank account or gift card balance, select it in the Amazon window first.

  • The approval gate stops the agent from buying through this server. It is not a sandbox: an agent with shell access to your machine could drive the browser's control port directly.

  • Automating your account may break Amazon's Conditions of Use. You are the one accessing your account, but Amazon can still act on it.

  • Amazon changes its pages. When a parser breaks, the tool says what it could not read instead of guessing.

Order history parsing uses amazon-orders (MIT).

Available Tools

8 tools
cartA
Read-only

What is in the cart now, with the subtotal.

Inputs: no named parameters. Returns: object

cart_addB

Add a product to the cart (one-time purchase, not Subscribe & Save). Returns the cart.

Inputs: asin (required), quantity. Returns: object

cart_removeB

Remove a product from the cart. Returns the cart.

Inputs: asin (required). Returns: object

checkout_requestA

Ask the user to approve buying the current cart. Opens Amazon's checkout review (nothing is placed), then an approval page in the user's browser showing total, items, address and card. Returns at once with status 'pending'; poll checkout_status. The user can approve or decline.

Inputs: no named parameters. Returns: object

checkout_statusA
Read-only

Where an approval stands: pending, switching (the user is changing the card), placing, placed (with order_number), rejected, expired, superseded or failed (with detail).

Inputs: approval_id (required). Returns: object

ordersA
Read-only

Order history with each item's return window. period: 'last30', 'months-3' or 'year-YYYY'. Amazon shows 10 orders per page; max_pages caps how many pages are read.

Inputs: period, max_pages. Returns: object

productA
Read-only

Title, price and availability for one product by ASIN (e.g. from orders).

Inputs: asin (required). Returns: object

returnable_itemsA
Read-only

Items from the last 3 months whose return window is still open, soonest deadline first.

Inputs: no named parameters. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 8 tools

    Disambiguation5/5

    Each tool has a distinct purpose: orders (history) vs returnable_items (open-window filter), product (lookup), cart/cart_add/cart_remove (view/add/remove), and checkout_request vs checkout_status (initiate vs poll). No two tools overlap in a way that would cause misselection.

    Naming Consistency4/5

    Mostly noun-based names with clear grouping: cart_ prefix for cart mutations and checkout_ prefix for the two-phase approval flow. Minor inconsistency in that orders is plural while product and cart are singular, but the convention is readable and predictable.

    Tool Count5/5

    Eight tools is well-scoped for a shopping/checkout assistant, covering the full browse-to-purchase flow without redundancy. Every tool earns its place.

    Completeness4/5

    Core lifecycle is covered: order history, returns, product info, full cart CRUD, and a two-phase checkout with status polling. Minor gaps exist (no product search beyond ASIN, no cart quantity update), but agents can work around these via existing tools.

    Related MCP Connectors

    Related MCP Servers