The day GitHub broke for 7 hours, Cursor launched its replacement. Coincidence? Sure. But 35% of merged PRs are now written by AI. Buckle up.
On the morning of August 17, 2026, Cursor began rolling out Origin, its own git hosting service, to everyone on a paid plan. Roughly three and a half hours later, GitHub's status page lit up with what became a nearly seven-hour global degradation: error rates around twenty percent on pull requests, issues, and the API, closer to fifty percent on archive and raw file downloads, with enterprise single sign-on and Copilot failing alongside. Matt Palmer, who works at Cursor, supplied the line of the day: "We were going to ship this earlier, but GitHub was down." A GitHub outage had delayed the launch of a GitHub competitor.
Nobody seriously believes Cursor timed it. Product launches get locked weeks in advance, and this was coincidence. But it was the kind of coincidence money cannot buy, because it acted out, in real time, the argument Origin exists to make: that where your code lives, for most of two decades the least interesting decision an engineering organization ever made, has become an interesting question again.
Stripped of the framing, what Origin actually is turns out to be straightforward. It is a git forge, hosted at origin.cursor.com, surfaced inside the Cursor editor as a new Codebase tab. You claim a codebase name, which becomes part of every repository URL, in the shape of cursor.com/codebase/acme-corp, create a repo, and push to it with ordinary git over HTTPS. A command-line tool called origin handles authentication and the surrounding chores. The service is in early beta, bundled into Cursor's paid plans with no separate price tag, no free tier, and no public repositories. Enterprise administrators can switch it off for their whole organization; notably, the default is on.
The product has a lineage worth knowing. In December 2025, Cursor acquired Graphite, the startup built around stacked pull requests and code review. Announcing that deal, Cursor wrote that "the boundary between where you write code and where you collaborate on it feels increasingly arbitrary" and teased "some more radical ideas we can't share just yet." Origin is the radical idea. Graphite's co-founder Tomas Reimers unveiled it on stage at Cursor's Compile conference in June 2026 and now leads its development. That same June day, as it happens, SpaceX signed its sixty-billion-dollar all-stock agreement to acquire Cursor's parent company, Anysphere. The deal closed on August 14, and Origin shipped the following Monday as the first product released by a SpaceX subsidiary. Reveal, close, launch: three separate dates that much of the coverage has collapsed into one, and shouldn't.
What shipped in the beta is deliberately conventional. Every repository gets pull requests with a timeline, commit history, status checks, and diff-level comments; you review and merge without ever opening a browser tab. There is code browsing and search. And there is a GitHub sync that amounts to the launch's most consequential design decision: connect a GitHub organization, pick repositories to mirror, and they appear alongside native ones, updating in real time, with pull request comments flowing in both directions within seconds. Pushes to synced repos keep landing on GitHub, which, in Cursor's own words, "stays the source of truth for anything started there." Asked on Hacker News what Origin does today that GitHub does not, Reimers was disarmingly candid:
"Today, very much nothing. We're intentionally releasing this as a GitHub alternative where we meet them toe-to-toe on functionality."
That candor undersells a few things. Poke at the CLI, as I did on the machine where I am writing this, and its footprint is larger than the changelog suggests: a full pull-request lifecycle with reviews, comment threads, and merges, plus snapshotted PR versions; mirrored-repo creation straight from GitHub; merge-time and push-time rulesets; SSH key management; and an origin api command modeled openly on gh api, pointed at a REST endpoint on api.cursor.com. Stacked pull requests, Graphite's inheritance, are built in, chained by a parent reference and mergeable as an entire stack, backed by a stack-aware merge queue. The agent story also shipped further along than a chat sidebar. Cursor's agents operate directly on hosted repositories: ask about the file you are browsing, hand the agent a review comment and have it revise the pull request in place, tell it to push a branch. Within days of launch, agents gained the ability to subscribe to their own pull requests, waking autonomously when a check fails or a reviewer comments and iterating until the PR is clean. The editor even ships internal skills that teach the agent to create and push repositories on its own initiative, which tells you who Cursor thinks the customer of this forge partly is: the software, not just its operators.
The integrations chosen for day one point the same way. Vercel gives every pull request a preview deployment and ships to production on merge. Depot and Buildkite provide continuous integration, and, in the strategically interesting part, both run existing GitHub Actions workflows unchanged. Cursor is not asking anyone to rewrite a build pipeline or retrain a team. It is asking them to open a second window onto code they already have, which is a far easier request to approve, and a classic wedge: the read-mostly mirror costs nothing to try and breaks nothing if abandoned, while quietly relocating the place developers spend their attention. If the review experience proves better, the system of record eventually follows it.
Honesty about the current state requires a long accounting of what is not there. No public repositories, so open source is out of scope entirely. No free tier. No issue tracker. No first-party CI, and for mirrored repos not even the partner CI, since Origin's apps only work on native repositories while Actions configuration and secrets stay behind on GitHub. Native Windows is unsupported; the CLI wants WSL. The rollout is staged, so even paying customers may not see it yet. And the paperwork a serious adopter needs, meaning pricing beyond "included," data retention, training-use policy, security architecture, and export tooling, remained unpublished at launch. VentureBeat rightly flagged that gap as the real open question, sharpened by the fact that the company asking to hold your repositories is now a wholly owned subsidiary of SpaceX, inside a division that also builds frontier models. For an enterprise, "opt-out, not opt-in" is a sentence worth reading twice.
So why take it seriously? Partly because of what Cursor says is coming, and partly because of the numbers behind why. The changelog's promise is blunt: "Agent-native features ship soon." Reimers put modest specifics on it in the launch thread: deeper agent integrations over the following weeks, tooling that helps make sense of agent-written code, and automation that moves pull requests toward a mergeable state on their own. He also described the conflict workflow Cursor already uses internally, in which an agent works in its own worktree, submits green work to a merge queue, the queue replays it on the latest target branch, and a conflicting change is rejected back to the agent to resolve and resubmit. "Agent-driven conflict resolution via a merge queue works really well in practice," he wrote. "It's almost necessary because of the increase in velocity of changes." Beyond that, the wishlist circulating among analysts, which includes provenance records for agent-authored commits, review surfaces organized around intent and verification evidence rather than raw diffs, and agents as first-class principals with scoped, revocable credentials, is exactly that: a wishlist. Cursor has committed to none of it publicly, and it is worth being plain that nobody, Cursor included, has shipped the agent-native forge yet.
The numbers, though, are the strongest part of the case. Pre-launch reporting surfaced the internal figure that best explains the product's existence: 35 percent of pull requests merged inside Cursor were opened by agents running autonomously in cloud VMs. GitHub's own Octoverse counted 43.2 million pull requests merged per month, up 23 percent year over year, while the big developer surveys keep finding the same asymmetry, with usage of AI tools climbing and trust in their output falling. Writing code stopped being the constraint; reviewing and integrating it became one. A forge built for humans assumes a pull request represents a person you can ask what they meant. Once a meaningful share of merged changes comes from software, the queue stops being a conversation and becomes a scheduling problem. That is a genuinely architectural argument for rebuilding the hosting layer rather than skinning it.
GitHub, meanwhile, has spent eighteen months making Cursor's argument easier. One analysis counted 257 incidents between May 2025 and April 2026, 48 of them major; Actions alone failed 57 times in a year; GitHub's CTO has said publicly that the platform must design for thirty times its current load. The Zig project decamped to Codeberg, Ghostty followed, and OpenAI reportedly began building an internal alternative after outages left its engineers unable to commit. GitHub's answer to the agent era, Agent HQ, orchestrates third-party agents on top of the existing platform, a coherent strategy that defends the substrate by conceding the layer above it. Origin attacks the substrate itself, with agents baked in rather than bolted on.
And yet the sober comparison still runs mostly GitHub's way. GitHub has the free tier, the open-source commons, first-party CI, issues, projects, a marketplace, and the enterprise apparatus of SAML, SCIM, audit logs, CODEOWNERS, and branch protections that Origin either lacks or has not documented. It has eighteen years of accumulated muscle memory and network effects that a beta bundled into one editor's paid plans cannot match, and an incumbent with the network graph can copy features faster than a challenger can build a network. Origin's own sync design concedes the point: for anything that started on GitHub, GitHub remains canonical, by Cursor's explicit choice.
Where does that leave someone deciding this week? Running both, almost certainly. Mirror a real repository, keep GitHub authoritative, and measure whether review inside the editor actually moves faster. The bet is reversible by construction, which is the most adoptable thing about it. The interesting version of Origin, the one that would justify the "GitHub killer" headlines, has not shipped; what shipped is a competent mirror with a very good review surface and a thesis attached. But the thesis is the right one. If agents keep writing a growing share of the world's code, and every measured trend says they will, then the forge that treats them as first-class users, with queues, provenance, and permissions designed for their tempo, is going to matter enormously. Cursor bought the right company, picked the right wedge, and shipped during the right outage. The next few changelog entries will tell us whether it can finish the argument.