The AI-native IDE that put a million developers inside an agent's workflow - and then became the most fought-over startup of 2025.
Open Windsurf today and something strange happens. The editor you installed last year has quietly become something else. In June 2026, an over-the-air update renamed it Devin Desktop. The default view is no longer a text editor at all - it is a command center, a Kanban board of AI agents writing code on your behalf, locally and in the cloud. Your settings carried over. Your muscle memory did not.
That update is the latest chapter in one of the odder corporate stories in software. Windsurf is a company whose founders work at Google, whose product belongs to Cognition, whose original agent - Cascade - was retired on July 1, 2026, and whose user base kept growing through all of it. Goldman Sachs writes code with it. So does Citi. So does Dell. So, depending on the week, do well over a million developers.
The tension that explains everything about Windsurf is this: software engineers spend most of their time not writing software. They read, search, refactor, test, deploy, and context-switch. Windsurf existed to collapse that overhead - and it did the job well enough that three of the biggest names in AI spent one weekend in July 2025 fighting over the pieces.
By 2022, AI code completion existed. GitHub Copilot could finish your line. What it could not do was understand your codebase - the half-million lines of context, the internal conventions, the dependency graph that makes enterprise software enterprise software. Suggestions without context are guesses, and guesses in a bank's codebase are incidents.
Windsurf's founders saw the gap from an unusual angle. They were not IDE people. They were infrastructure people - their first company, Exafunction, virtualized GPUs for large-scale machine learning workloads. When large language models arrived, they realized the inference stack they had built for other people's models could serve their own. The problem worth solving was not "predict the next token." It was "make a model genuinely useful inside a real, messy, regulated codebase."
That framing - context first, model second - became the company's whole identity. It is why Windsurf built its own codebase-indexing engine, trained its own SWE model family for software engineering tasks, and chased security certifications most startups would not touch with someone else's compliance team.
Varun Mohan and Douglas Chen met in middle school, then again at MIT. In 2021 they founded Exafunction and raised from Greenoaks and Founders Fund to do GPU virtualization. It worked. Then Copilot launched, and they made the kind of decision that sounds easy in retrospect and feels like vertigo at the time: they walked away from a working business to build a free code autocomplete tool called Codeium.
Free was the strategy, not a promo. Unlimited autocomplete, 70+ languages, plugins for 40+ editors. The free tier was a land grab for developers and - usefully - a firehose of feedback for the models. By early 2024 Kleiner Perkins led a $65M Series B; by September 2024 General Catalyst led a $150M Series C at a $1.3B valuation, bringing total funding to roughly $243M.
The second pivot was the bigger one. In November 2024 the company shipped the Windsurf Editor, an AI-native IDE built around an agent called Cascade - software that could plan and execute multi-step, multi-file changes with a human approving the work. Plugins ride inside someone else's editor; Windsurf decided to own the whole surface. Five months later the company renamed itself after the product. The tool had become the company, which is its own kind of confession.
Mohan and Chen start a GPU virtualization company backed by Greenoaks and Founders Fund.
Copilot launches; the team repurposes its inference stack into a free AI autocomplete tool.
General Catalyst leads; total funding reaches ~$243M.
One of the first agentic AI-native IDEs, built around the Cascade agent.
The company renames itself after its flagship editor.
Exclusivity on OpenAI's offer expires. Google pays ~$2.4B to license the tech and hire the founders plus ~40 R&D staff.
Product, brand, IP, ~210 employees, $82M ARR, 350+ enterprise customers. First call Friday evening, signed by Monday morning.
Agent Command Center becomes the default view; Rust-based Devin Local succeeds Cascade; open Agent Client Protocol support lands.
Strip away the deal drama and Windsurf is a coherent set of tools aimed at one outcome: keeping developers in flow while agents handle the repetitive middle of software work.
Windsurf Editor / Devin Desktop. An AI-native IDE on a VS Code foundation. Since June 2026, the default view is the Agent Command Center - a dashboard for running and reviewing multiple coding agents, local and cloud, in parallel.
Cascade → Devin Local. The original agent planned and executed multi-step, multi-file changes with human approval, grounded in a proprietary codebase index. Its successor, Devin Local, is rebuilt in Rust with subagent support and reportedly up to 30% better token efficiency.
Codeium extensions. Autocomplete, chat, and search across 40+ editors and 70+ languages - the famously generous free tier that built the user base.
Windsurf Enterprise FedRAMP. Via Palantir FedStart and AWS GovCloud: FedRAMP High, DoD IL4/IL5, and ITAR support, with code data kept transient and encrypted rather than stored.
The SWE model family. Proprietary models (SWE-1, SWE-1.5) trained specifically for software engineering workflows, alongside access to frontier models including the latest Claude models.
For a developer, the practical pitch is blunt: describe the change in plain language, let the agent draft it across the codebase, review the diff, ship. For an enterprise, the pitch is the part competitors found hardest to copy - the certifications. A regulated bank or a defense contractor cannot use most AI tools at all. Windsurf made itself one of the few they could.
Numbers first. At the time of the July 2025 split, Windsurf reported $82M in annual recurring revenue, with enterprise ARR doubling quarter over quarter, 350+ enterprise customers, and hundreds of thousands of daily active users. Those numbers explain the bidding war better than any press release.
OpenAI agreed to buy the company for roughly $3B - then the deal collapsed as its exclusivity expired, with Microsoft's partnership terms reportedly a complicating factor. Within hours, Google signed a ~$2.4B arrangement: a non-exclusive technology license plus the hiring of Mohan, Chen, and about 40 senior researchers into DeepMind. Reports later indicated roughly half went to investors and half to compensation for the hired team.
That left a company with a product, customers, revenue - and no founders. Cognition, maker of the Devin coding agent, made its first call after 5 PM that Friday and signed a definitive agreement by Monday morning, reportedly for about $250M. One detail traveled further than the price: Cognition accelerated equity and waived vesting cliffs so every remaining employee got paid - a pointed correction to a deal that had left most of the staff behind.
Today the customer roster runs from solo developers on the free tier to Goldman Sachs, Citi, and Dell, with 2026 reports citing thousands of enterprise customers across the combined Cognition portfolio. Inside Cognition - which reportedly raised at a ~$25B valuation in 2026 with Windsurf integrated and ARR doubled - the IDE anchors the desktop end of the Devin agent ecosystem.
Windsurf's stated aim was always to accelerate software engineering by keeping developers in flow - grounding AI in real codebase context and letting agents handle multi-step work while humans keep approval rights. Under Cognition, the framing has sharpened: developers stop typing most of the code and start managing the agents that do. The Agent Command Center is that idea turned into furniture - the kanban board where you watch your fleet work.
It is also, notably, an open bet. Devin Desktop ships with support for the Agent Client Protocol, meaning rival agents - Codex, Claude Agent, OpenCode - can run inside it. For a product born of a bidding war between competitors, hosting the competition is either confidence or excellent comedy. Possibly both.
Return to that June 2026 update - the morning Windsurf woke up as Devin Desktop. A year earlier, the same product was a text editor with a clever agent in the sidebar, owned by a startup with two founders and a $3B offer on the table. Now the editor is the sidebar. The agents are the product. The founders are at Google, the company is part of Cognition, and the developers - the only constituency that never signed a term sheet - got a tool that kept improving through all of it.
That is the real measure of Windsurf. Not the four price tags, not the 72-hour deal, but the fact that the central tension it was built around - engineers drowning in everything except engineering - is being resolved in public, one release at a time. The scene from the opening has changed: the developer is no longer typing. She is reviewing, directing, deciding. The typing is delegated.
Whether that future belongs to Devin Desktop or to one of its rivals is an open question. But the company that did the most to force the question fit it all into five years, three names, and one unforgettable weekend. Worth watching what it does with year six.