Menlo Ventures led a $400 million Series C at a $13.3 billion valuation [8], against a run rate that Menlo partner Deedy Das puts above $500 million annualized [7]. That works out at roughly 27 times revenue [13], which is not the price of a code generator with visible competitors. Lovable's own account of its market puts it in the same water as Vercel and Cloudflare, and notes that Vercel's v0 already turns natural language into an app [10].
The growth numbers point the same way. More than 60 million projects created against more than 900 million monthly visits to Lovable-built apps [7] averages about 15 visits per project per month [14]. Most of that inventory is not an audience. It is functions sitting behind a login, which is the shape of thing worth calling from somewhere else.
Hedin's account of how the company got here explains the target. Users moved from prototypes to production products and then to the software that runs the company, including admin panels and customer-support consoles [6]. That is where the write actions live, and a capability by Lovable's own definition is a function an agent invokes with nobody opening the app [2].
Which makes one word the load-bearing part of the pitch: selected. Lovable exposes selected functions from a published app as tools through a hosted MCP server [3]. The Latent Space account does not say who authorises a call, how a capability is scoped, or whether calls are logged [15]. For an internal console at a company where, by Lovable's own claim, employees at nearly two-thirds of the Fortune 500 have used the platform [9], those are the questions that decide whether the feature ships to production or stays in a sandbox.
Hedin's framing is that application-layer releases and improving models compound on each other [11], and that the destination is a single entry point to all your work [4] in which everything you build is reusable in an agentic way [5]. Read as a product plan rather than a vision, that demotes the human UI to the optional half of a delivery. The company that hosts the app and the tool server also becomes the thing that describes your software to every model that asks. Getting that description wrong is not a rendering bug; it is a function called on real data by a client you do not administer.
