Most people running commerce operations already spend a large share of their day inside Claude. They draft emails in it, analyze spreadsheets in it, prepare for meetings in it, and increasingly ask it questions they used to answer by logging into three different dashboards. The interesting question is no longer whether AI belongs in the workday, but why the rest of the work still requires logging into a half dozen separate applications. With the recent advancements in Claude and the tooling around it, that requirement is starting to disappear, and it’s now realistic to run entire commerce operations headless, with no application UI in sight.
Twenty years of walled gardens
SaaS architecture has looked more or less the same for twenty years: a vendor ships a UI plus an API plus an application plus a database. Data goes in, gets transformed, and comes back out. Each platform is a walled garden with its own login, its own UI, its own app marketplace, and its own integration layer. Salesforce, ServiceNow, Shopify, and a hundred others all follow the same blueprint, and they are all very much closed systems.
The result is familiar to anyone who has worked in commerce operations. The same business user logs into three, five, or a dozen tools, each with its own interface and its own fragmented slice of the data. Nothing shares a spine, so connecting any two of them is a project rather than a given. In that world, applications differentiated on features, more features meant the vendor sold more, and that math drove every roadmap in the industry. The cost landed on the customer, who took on more implementation work and more complexity with every feature added. For two decades that tradeoff held because there was no alternative.
Apps are becoming capabilities
There is now an alternative. Claude has become the new super app for business, complete with its own interface, its own integration layer in MCP, and its own marketplace of tools and skills. In this world, applications stop being destinations and start being capabilities – things an AI reaches for over MCP when the work calls for them, the same way a person reaches for a calculator.
Nicolas Dessaigne, co-founder of Algolia, put it well: “The new moat in the agent era is being the tool agents reach for.” A coding agent doesn’t reinvent a database, it wires up Supabase. The same pattern is repeating across every category. Legal databases that used to force users into a portal now surface caselaw directly into Claude over MCP. Development platforms are shipping agent-first surfaces alongside their human ones. The vendors that win will be the ones easiest for an agent to find, reason about, and wire up.
This shift also explains why features are commoditizing so quickly. Anyone can now vibe code a feature on demand, so feature depth stops being a moat almost overnight. When every vendor in a category can reach feature parity for free, differentiation has to come from somewhere else – from networks, from proprietary data, from deep vertical expertise, and from being the thing AI reflexively reaches for.
What running headless actually looks like
Consider what business users in commerce actually do all day. A VIP segment is churning. Returns on a new line of jeans are climbing. One of the 3PLs is backed up and taking 36 hours to ship orders. The Q2 margin target is off by 4%. Solving any one of these used to mean logging into the commerce platform, the OMS, the ERP, and the CRM, each with a separate login and a separate approach to workflows and approvals.
Running headless means the operator stays in Claude. They ask which 3PL is backed up, Claude pulls live order and fulfillment data through MCP, and the reroute happens through the same connection. Exceptions get investigated conversationally rather than through a queue of screens. Where a screen is genuinely needed, MCP Apps render small pieces of interface directly inside the conversation rather than sending the user off to a full application. The standards have matured to match: MCP went stateless in the mid-2026 release, which makes it dramatically easier to surface inventory, shipping, and orders directly to any front end, and the Order Network eXchange (onX) now handles the commerce-specific post-purchase operations alongside it.
The application doesn’t disappear in this model. It moves back one layer, from a destination the user visits to a capability the AI consumes.
Pipe17 has been ahead of this curve
Pipe17 was the first order management vendor to ship full MCP support, launching in June 2026, and the bet behind that launch is now paying off. With the latest advancements in Claude, running Pipe17 headless has gone from a demo to a realistic operating model, and Pipe17 customers are starting to do exactly that – managing orders, inventory, and fulfillment across their selling channels, ERPs, and 3PLs from inside Claude, without opening the Pipe17 dashboard.
Every capability in the platform is exposed the same way. Order routing, inventory sync, exception handling, and fulfillment orchestration are all available to Claude as tools, and over time customers can mix and match those capabilities with their own, building exactly what they need on top of the network instead of waiting on anyone’s roadmap. The walled garden is gone; the capabilities remain.
Horizontal intelligence, vertical expertise
The division of labor here matters, because neither side can do the other’s job. Claude provides the horizontal AI expertise: general reasoning, language, planning, and the judgment to work across every domain a business touches. What the AI labs will not do is learn the messy particulars of how commerce operations actually run – the 3PL that behaves differently than its documentation says it will, the SKU mapping that broke when an API changed, the bundle that partially allocates, the warehouse that accepts an order it later can’t ship.
That vertical expertise is what Pipe17 supplies, and it rests on something AI can’t generate: a network of hundreds of live connections to selling channels, back-office systems of record, and fulfillment providers. Those connections represent years of accumulated knowledge about how real systems behave in production, and every order that moves through the network deepens that knowledge in a way the next customer inherits automatically. Twilio wasn’t valuable because of its dashboard, it was valuable because it made the telecom network programmable. Stripe made payments programmable, and Plaid made banking programmable. Pipe17 does the same for commerce operations, and in an AI-native world that network becomes the way Claude reaches the physical world of orders, inventory, and fulfillment.
A better model only makes that pairing stronger. Claude gets smarter at reasoning; the network keeps getting deeper at the things reasoning alone can’t reach.
The head does the thinking, the network does the work
The walled gardens will fade as destinations. Business users have already voted with their attention, and the work is consolidating inside Claude faster than most vendors want to admit. The vendors that thrive in what comes next won’t be the ones with the longest feature lists, they’ll be the ones that made themselves the easiest thing for an agent to reach for.
That’s the role Pipe17 has chosen. Claude brings the intelligence, Pipe17 brings the connectivity and the commerce expertise, and the operator gets something neither could deliver alone: commerce operations that run from a single conversation. Pipe17 helps Claude understand commerce.
