The Database That Learned to Disappear

In 2021, three people who had helped Dropbox move an ocean of files off Amazon's cloud found themselves staring at the same consoles everyone else sees: queues, virtual networks, databases, routing tables, permissions. Jamie Turner, James Cowling and Sujay Jayakar knew this machinery unusually well. That was precisely the problem. They knew how much of it had nothing to do with the app they wanted to make.

Their complaint was not that cloud infrastructure was weak. It was that it spoke the language of infrastructure teams. A small product company merely inherited the organizational chart of a very large one: storage over here, networking over there, an alert at 2 a.m. somewhere in the middle. So they paused their prototypes, talked to other startups and formed Convex around a mischievous premise. What if the backend behaved less like a construction site and more like part of the programming language?

“We just wanted to build products.”Convex's 2021 launch essay

The product is the missing glue

Convex is often called a database, which is accurate in the way a kitchen is accurately called a cupboard. The managed platform stores data, but it also runs server-side TypeScript functions, pushes query results to connected clients, caches them, schedules work, stores files, searches text and vectors, and talks to outside services. A React application can subscribe to a query. When a dependency changes, Convex reruns the query and sends the new result. The developer does not assemble a WebSocket service, cache policy and invalidation scheme to make the screen current.

This is the distinction from a modular Postgres stack. Supabase, Firebase, AWS Amplify and a hand-picked combination of Neon, functions, storage and messaging can all produce capable applications. Convex's argument is narrower and more opinionated: the seams themselves are a tax. Its query and mutation functions are transactions. Its type information runs from schema to client. Its cache and subscriptions understand the same dependencies. You trade some familiarity and direct control for fewer opportunities to make the pieces disagree.

Then the customer acquired a robot colleague

For years, Convex encountered a predictable objection: it was not Postgres. Experienced engineers had accumulated scar tissue, libraries and strong opinions around conventional databases. Asking them to adopt a new data model was asking them to abandon useful expertise.

Coding agents changed the sales conversation. If a machine writes much of the implementation, the human cares less about whether it used a beloved query builder and more about whether the result survives concurrent users. Convex has leaned hard into this opening. Serializable transactions retry on conflict. A made-up field name can become a TypeScript error. Components isolate their own schema, functions and data behind an explicit API. The company describes a component as small enough to fit inside an agent's context window - a wonderfully current unit of architecture.

The numbers are company-reported, but the direction is plain. Convex moved from frontend developers and small startups toward businesses that ask for SSO, audit logs, dedicated capacity and contracts. Named enterprise customers include Conduit, doxy.me, Orchid, Reducto and Town. A platform that once promised an afternoon prototype now also promises the paperwork required to keep it in production.

The useful failure came first

Convex's cleanest lesson is not in the launch material. It is in an admission. The company released its backend source in March 2024: more than 200,000 lines, largely Rust, with the same core code used by its cloud service. But the dashboard was missing. Data storage was limited to SQLite. Installation meant compiling and managing the system yourself. A year later, CTO James Cowling wrote that the release was real but “wasn't particularly useful.”

The license remains a compromise. Convex uses the Functional Source License, which allows broad use but prevents a direct competing service; each release converts to Apache 2.0 after two years. And the company still recommends its managed cloud for scaling, backups and high availability. The escape hatch is real, not effortless. Anyone self-hosting inherits the operational work that Convex was founded to remove.

“Code is cheap now. Design is not.”Jamie Turner, announcing the 2026 Series B

What it costs - and what you are buying

The commercial model is a familiar staircase. A free tier and pay-as-you-go Starter plan court experiments. Professional costs $25 per developer each month, with usage allowances and lower unit rates. Business begins at a $2,500 monthly minimum. Enterprise is negotiated. Convex also charges for resources such as calls, compute, storage, search and data egress once included amounts are exhausted.

The money is not purchasing a database in isolation. It purchases the absence of a backend team assembling and watching several services. That bargain improves when the application is collaborative, live-updating and TypeScript-heavy. It weakens when a team needs direct SQL, an unsupported server language, deeply mature offline behavior or ad hoc analytics in the primary store. Convex itself conceded in 2025 that ad hoc analysis “frankly” lagged and named OLAP support as work to do. High-traffic customers can also face substantial usage bills; the company has acknowledged some were paying more than $10,000 a month before its enterprise tiers arrived.

The playbook hidden in the plumbing

There are three things another company can copy without building a database. First, sell a reduction in decisions, not a multiplication of features. Convex lists many capabilities, but the emotional product is one model that stays coherent. Second, put the safe path beneath the user's feet. A transaction that retries by default is more valuable than a page explaining how to add the right lock. Third, audit the escape hatch as a product. Source code that technically exists is not the same as a system customers can actually leave with.

Convex applies the same opinionated quality to its culture. Its five published values include “Be Worthy of Trust,” “Make Work Human” and the unusually unvarnished “Give a Shit.” Each comes with intentional costs. In-person work may be noisy. Diversity can slow decisions. People are expected to disagree without asking the company to settle every question. The point is not that these are universal virtues. It is that a value without an uncomfortable tradeoff is decoration.

The company now sits in an interesting place. It is a developer tool selling to people who increasingly delegate tool choice. It is an infrastructure company arguing that infrastructure should be less visible. And it is an opinionated platform offering self-hosting because customers reasonably fear opinions they cannot escape.

That tension is the story. Convex does not make the backend disappear. Databases, transactions, storage engines and failure modes remain, tended by people who think about them all day. It simply changes who must notice. For a certain kind of builder - especially one working beside an eager coding agent - not noticing may be the feature worth paying for.