Saoud Rizwan Built a Coding Agent That Makes Nothing On Your AI Bill
He wrote the first version of Cline ten days after Claude 3.5 Sonnet shipped, entered it in a hackathon, and lost. Two and a half million installs later, the open-source coding agent is trusted by Samsung, SAP and Oracle - and its founder still refuses to mark up a single token.
In June 2024, ten days after Anthropic released Claude 3.5 Sonnet, Saoud Rizwan sat down to read the model card. Most people skimmed the benchmarks. He kept going, to a quieter part about agentic coding - the model's ability to work through a task step by step rather than autocompleting a line. He noticed the tools everyone was using, GitHub Copilot and Cursor, weren't really built for that yet. So he opened an editor and started building the thing that was missing.
He wrapped it into a submission for Anthropic's Build with Claude hackathon. It did not win. That detail matters less than it sounds, because the project - then called Claude Dev, later Cline - had already started doing the only thing that counts: developers who tried it kept it open. By October 2024 he launched it publicly. By the summer of 2025 it had passed 2.7 million installs and raised $32 million.
I realized I opened up this can of worms.Saoud Rizwan, on the moment Cline started to catch on
01 / The SetupIndiana to a hackathon that mattered
Rizwan grew up in Indiana and made his way to San Francisco. He is 28. Before anyone was pitching him, before enterprise logos, he was the kind of engineer who ships small, useful things and gives them away. His GitHub history is a run of open-source Swift libraries with plain, almost stubbornly literal names: Disk for saving files on iOS, Dance for animation, Piano for sound and haptics, CardSlider for the swipe-card interface. None of them were trying to be a company. They were tools he wanted to exist.
That instinct - build the tool you wish you had, then hand it to everyone - is the through-line. When Claude 3.5 Sonnet arrived with real coding ability, he pointed the same reflex at a much larger surface: the daily work of writing software itself.
02 / The IdeaAn extension, not a fork
A lot of AI coding startups made the same early bet: fork VS Code, ship your own editor, own the whole surface. Rizwan went the other way. Cline is an extension that lives inside the editor you already use. His reasoning was practical rather than ideological - Microsoft ships updates to VS Code fast enough that maintaining a fork becomes a treadmill of broken APIs. An extension is harder to make flashy, easier to keep alive.
Inside that extension, the product philosophy is almost aggressively minimal. He resists adding controls for their own sake.
Just keep it really simple. Just let the developer feel like they can use it however they want.On product design
The signature feature, Plan and Act, wasn't dreamed up in a roadmap meeting. Rizwan watched what developers were already doing: before letting the agent touch anything, they'd write out a markdown plan of the architecture, almost like briefing a coworker. So he built that behavior into the product. In Plan mode the agent is exploratory, asking questions and mapping the codebase; in Act mode it executes. The idea was sitting in front of him, in the way people already worked.
Fig. 1 - Cline traction since the October 2024 launch
Directional timeline of adoption. Cline crossed 2.7M installs across the VS Marketplace and Open VSX Registry, alongside 65k+ GitHub stars.
03 / The Contrarian BetZero margin on inference
Here is where Rizwan gets genuinely unusual for the moment we're in. Most AI products quietly buy tokens wholesale and sell them retail - a margin baked into every request you make. Cline doesn't. Developers pay the model providers (Anthropic, Google, OpenAI) directly, and Cline shows a full breakdown of what each request cost. The company takes nothing off the top.
We capture zero margin on AI usage. We're purely just directing inference.On the business model
He states the principle even more flatly: "Inference cannot be the business model." The bet is that transparency compounds into trust, and trust is the thing you can actually build a company on. Because Cline is open source, a skeptical engineer can read the code and see exactly how the agent talks to the model - no hidden prompts, no mystery markup.
It comes with a trade-off he's happy to name out loud. Cline doesn't cut corners with retrieval tricks to save tokens; it throws full context at frontier models. That produces better results and a real bill.
You kind of get what you pay for. Throw a task in Cline and it gets expensive.On why Cline doesn't cut inference costs
04 / The TurnFrom hobbyists to Fortune 500
Cline started, in Rizwan's own framing, as an open-source project for hobbyists. What happened next surprised the ordering that most startups expect: the individual developers came first, and the enterprises followed them in. The same properties that won over a solo engineer - transparency, control over models, no data games - turned out to be exactly what a security team at a large company wants before it lets an AI near its codebase.
By 2026, Cline counts Oracle, Samsung, SAP and Salesforce among the organizations using it, and it powers Oracle's NetSuite SuiteCloud Developer Assistant. The pitch to those teams is the same one that hooked the hobbyists: full control over your code, your models, and your costs.
05 / The TeamHire people who like making life easier
Rizwan tends to hire from his network - people who already understand why the problem is exciting. His read on what makes coding-agent work compelling is refreshingly unglamorous.
Engineers in general love working on things that make their own lives easier.On building the Cline team
It's a small idea that explains a lot. Cline isn't a tool built for some abstract user; it's built by engineers scratching their own itch, which is why the product feels like it was designed by someone who actually writes code all day. The company backing it - roughly three dozen people as of this writing - grew out of that same instinct, not out of a market-sizing spreadsheet.
06 / What He's AfterSafe AI, inside the enterprise
The larger ambition is to bring agentic AI into serious software organizations without asking them to give up the things they can't give up: control of their code, choice of their models, visibility into their spend. Rizwan has become one of the more credible voices arguing that you can move fast on AI and stay in control at the same time - that those aren't opposites.
There's a tidy irony in the arc. A project that lost a hackathon judged by experts went on to be adopted by millions of developers and some of the largest companies on earth. Judges pick winners. Users make them. Rizwan mostly kept his head down and let the second group vote.