A product that now accounts for roughly 4% of all public GitHub commits began as a VS Code extension that its own team “basically forgot about” for most of 2022. Anthropic published that origin story last week, and the trajectory it describes \u2014 from safety-lab experiment to the fastest revenue ramp in enterprise software history \u2014 says more about where AI development tools are heading than any benchmark score.<\/p>\n
As of this week, Claude Code continues to ship updates at a pace that matches its reputation: the July 14 changelog adds screen reader support, enterprise process-wrapping infrastructure, and a raft of bug fixes \u2014 the product’s relentless release cadence continuing one week after Anthropic dropped a sweeping oral history of how it was built.<\/p>\n
The oral history of its development<\/a> draws on recollections from co-founders and more than a dozen engineers and researchers. What it reveals is a development arc defined less by strategy than by accident, persistence, and one engineer’s insistence on building for models that didn’t exist yet.<\/p>\n Anthropic’s First Product Was a Coding Tool Nobody Planned<\/p>\n When Dario and Daniela Amodei left OpenAI in 2021 with a dozen researchers to found Anthropic<\/a>, the company’s stated purpose was AI safety research \u2014 not shipping products. The decision to build anything at all was, in co-founder Ben Mann’s words, “a fairly controversial decision at the time.”<\/p>\n When the team resolved that debate, their first product was not a general assistant. It was a coding tool: a VS Code extension that offered users four different code suggestions per prompt. Research engineer Dawn Drain, who had just joined the company, set herself a deceptively modest goal: get the model’s coding ability to at least match her own.<\/p>\n Simultaneously, reinforcement learning lead Shauna Kravec was pursuing a more ambitious vision. Her team wasn’t interested in autocomplete. They wanted a model that could “do work” \u2014 specifically, autonomous software engineering. Their entry point: if AI could eventually handle complex knowledge work, it would have to master software engineering first. They spent 2022 training models with reinforcement learning to write and verify simple functions. The early results were, by Kravec’s own account, “really terrible.”<\/p>\n The coding assistant briefly went live with roughly 100 external users before the team pivoted to shipping the Anthropic API. It was, as Mann put it, “basically forgotten about for a while.”<\/p>\n How Agentic Coding Works \u2014 and Why the Technical Leap Mattered<\/p>\n The technical foundation that eventually became Claude Code emerged from two parallel advances. By late 2022, Kravec’s team had given the model Bash tool-calling capabilities \u2014 the ability to freely search and execute commands within codebases. Dawn Drain spent what she called “an embarrassingly long time” teaching Claude to write code diffs, the most natural format for modifying existing code.<\/p>\n These advances converged in an internal command-line tool called clide, coined by engineer Eli Tran-Johnson. Clide’s key architectural innovation: it could summon up to 100 Claude Haiku models in parallel to read and answer questions across a codebase too large for any single context window. Where GitHub Copilot suggested the next line of code, clide could understand an entire repository and propose changes across dozens of files simultaneously.<\/p>\n The resulting product architecture \u2014 what researchers now call an “agentic loop” \u2014 works fundamentally differently from autocomplete. The agent receives a high-level prompt, evaluates the current state of the codebase, decides what tool to call (read a file, run a test, execute a shell command), receives the result, and repeats until the task is complete. Claude Code lead engineer Adam Wolff added early “agent capabilities” \u2014 the ability to infer user intent from partial changes. The first time this worked, Wolff recalled, he was “dancing in the kitchen.”<\/p>\n This architectural difference is why Claude Code’s 46% “most loved” rating in The Pragmatic Engineer’s February 2026 survey of 906 developers<\/a> so dramatically exceeds GitHub Copilot’s 9%: the tools are doing categorically different things. Copilot is fast at line-level suggestions. Claude Code can complete a feature.<\/p>\n The model-capability threshold matters here. Dawn Drain put it plainly: “Once you cross the model capability threshold, the product form reveals itself.” Clide existed years before that threshold was crossed. It was, by one creator’s assessment, “wonky” and “very, very ahead of its time.”<\/p>\n Safety Research as Competitive Advantage<\/p>\n The most underreported aspect of Claude Code’s success is the causal link between Anthropic’s AI safety mission and its enterprise dominance. The draft’s “from safety lab to Silicon Valley” framing implies tension \u2014 a reluctant commercial pivot. The actual story is more interesting.<\/p>\n Anthropic’s safety-first culture produced the technical discipline \u2014 reinforcement learning with human feedback, systematic evaluation frameworks, careful capability assessment \u2014 that enabled Claude Code’s agentic architecture to function reliably enough for enterprise adoption. The same research culture that initially made “building products” a controversial internal debate created the evaluation rigor that distinguishes Claude from tools that hallucinate confidently.<\/p>\n Eight of the ten Fortune 10 companies now use Claude, a fact confirmed in Anthropic’s Series G announcement<\/a>. The adoption pattern is not coincidental. Enterprise buyers choosing to automate software engineering workflows need to believe the tool won’t delete the wrong database. The trust comes from somewhere.<\/p>\n The Two-Week Sprint That Built Everything<\/p>\n In January 2024, Mann established Anthropic’s Labs team, explicitly targeting what he saw as a gap in the market for “agentic coding.” That September, Boris Cherny joined with a mandate for “agentic programming” and a piece of advice that would define the project: “Don’t build for today’s models. Build for the models six months from now.”<\/p>\n Cherny’s first prototype \u2014 assembled in two days \u2014 was a minimal CLI that could screenshot Apple Music and identify the currently playing song. He posted it to Anthropic’s internal Slack and received two or three likes. Nobody, including Cherny himself, fully understood what they were looking at.<\/p>\n What changed his mind was clide. When Adam Wolff rejected one of Cherny’s pull requests and told him to try the internal tool instead, Cherny pasted in the issue description. Clide wrote the complete five-to-ten-line fix. “From that moment,” Cherny recalled, “an unprecedented sense of urgency seized him.”<\/p>\n In December 2024, the Labs team launched a final push. A team of two \u2014 Cherny and engineer Sid Bidasaria \u2014 swelled to eight or nine engineers. In two weeks, they built nearly everything that shipped: bug reporting, login flow, auto-updates, usage metrics. No PR review restrictions. Fixes shipped immediately. Users could have a new version within five minutes.<\/p>\n The small team size was, counterintuitively, a feature. Understaffed, these engineers were forced to use Claude Code obsessively to keep pace. By the time the product launched in February 2025, approximately 90% of its codebase had been written by Claude Code itself \u2014 a claim that is both a technical milestone and a proof point that the team understood well.<\/p>\n Mann captured the product philosophy: “You have to now build products that only have a 20% to 30% success rate, so that when the next model comes along, its success rate can reach 80%.”<\/p>\n How Claude Code Democratizes Who Gets to Build Software<\/p>\n The oral history’s most humanizing passage describes a nonprofit in Alaska that works with disabled teenagers. Before Claude Code, the organization coordinated activities with pen and paper. With it, they built a custom tablet application and automated the transcription of handwritten fuel delivery logs into a structured database \u2014 work they could never have afforded to commission.<\/p>\n This is the dimension of Claude Code’s impact that revenue figures don’t capture. Product lead Cat Wu observed a quieter behavioral shift: at launch, users anxiously read every permission request the tool initiated. By mid-2025, most had switched to selecting “Auto-accept all.” Trust, the team learned, accrues \u2014 it isn’t a feature you ship.<\/p>\n Cherny has a personal timeline for this transition. In February 2025, Claude Code helped him write about 10% of his code. By May, 40%. By winter 2025, 100%. His current day can include up to 88 commits, not a line written by hand.<\/p>\n What the Record Revenue Numbers Actually Measure<\/p>\n The growth metrics deserve careful framing. Claude Code reached $1 billion in annualized revenue within six months of its May 2025 general availability launch \u2014 a rate multiple analysts describe as the fastest $1B ramp in enterprise software history, though the comparison is based on run-rate revenue rather than recognized revenue, and the “fastest ever” superlative is asserted primarily by industry analysts rather than by Anthropic itself, as confirmed in Anthropic’s own Series G disclosures<\/a>.<\/p>\n By February 2026, that figure had grown to $2.5 billion, confirmed by Anthropic’s Series G announcement<\/a> alongside Reuters’ coverage of Anthropic’s $30 billion funding round at a $380 billion valuation. As of February 2026, SemiAnalysis’s February 2026 analysis<\/a> estimated that Claude Code was authoring approximately 4% of all public GitHub commits \u2014 roughly 135,000 per day \u2014 with a projection that the share could exceed 20% by the end of 2026.<\/p>\n