Cursor Origin and the End of the PR Page as Review Surface

On August 17, 2026, Cursor began rolling out Origin to paid users. Three and a half hours later, GitHub degraded for six hours and forty-two minutes, taking down pull requests, the API, archive downloads, enterprise SSO, and Copilot simultaneously. The two events were unrelated. The fact that they shared a news cycle made the argument for an alternative git host more vivid than any product launch could have managed. But the outage is the smaller story. The deeper one is what happens to code review when the tool that writes the code also hosts it.

What Origin actually is#

Origin is a forge: the service layer around Git that handles storage, permissions, pull requests, checks, and merges. It lives inside a new Codebase tab in Cursor. Teams push to it over the command line and get a full pull-request surface inside the editor, including timelines, diffs, reviewer comments, and merges, without loading a browser.

The design decision that matters most for enterprise adoption is the mirror mode. GitHub stays the source of truth for any repo started there. Pushes go to GitHub; Origin syncs pull-request conversations in both directions within seconds. That architecture makes Origin safe to evaluate without touching a single branch protection rule. It also makes it easy to adopt incrementally and easy to abandon if the terms eventually turn unfavorable.

Day-one integrations shipped with Vercel, Buildkite, and Depot. Existing GitHub Actions workflows run unchanged on Buildkite. A forge that runs your existing CI and deploys previews to the CDN you already pay for is a candidate for real work. A forge that requires rewriting pipelines is not, and Cursor was careful not to be the latter.

The 35% number that explains the timing#

Michael Truell, Cursor's co-founder and CEO, reported in February that 35% of pull requests merged inside Cursor were created by agents running autonomously in cloud virtual machines. That is a Cursor-internal metric, not an audited figure. But take it at face value and the implication is structural: a forge designed for human authors fails when a third of its input stops being human.

A pull request from a person carries implicit context. You can ask what they meant. An agent PR carries a task description and a diff. The queue becomes a scheduling problem, not a conversation. Cursor's demonstration at the Compile conference claimed throughput of 22.6 commits per second per repository; the company has not published a benchmark methodology for that figure. The point, though, is architectural: the bottleneck moved. Generation is fast. Review and integration are not.

Google's 2025 DORA report, drawn from nearly 5,000 technology professionals, found that 90% of developers use AI coding tools, and that AI adoption correlated positively with delivery throughput and negatively with delivery stability. More output, more breakage. Stack Overflow's 2025 developer survey of 49,009 respondents found that trust in AI accuracy fell to 33% from 43% in a single year, while distrust climbed to 46% from 31%. The volume grows anyway.

GitHub's reliability record made this argument easier to make#

The August 17 outage was the seventh incident on GitHub's status page in fifteen days. An analysis by LeadDev counted 257 incidents between May 2025 and April 2026, 48 of them major, roughly one significant disruption per week. GitHub Actions alone accounted for 57 outages across those twelve months. GitHub CTO Vlad Fedorov has said the platform was not built for its current scale and must be designed for 30 times today's load.

The defections preceded Origin. The Zig programming language moved to Codeberg in November 2025, citing Actions failures. Mitchell Hashimoto announced in April that Ghostty, a terminal emulator with more than 52,000 stars, would leave too, pointing to near-daily outages blocking reviews and CI. The Information reported in March that OpenAI began building its own GitHub alternative partly because outages left engineers unable to commit for hours at a time. Enterprises that assumed GitHub monoculture was the only sensible default are now revisiting that assumption, and multi-forge review tooling has moved from a niche concern to a procurement question.

The review layer cannot live on the same surface as the code#

Here is where the structural problem sharpens. When the editor, the git host, and the model all belong to the same vendor, the pull request page is no longer a neutral surface. It is a product feature of the company whose agent wrote the change being reviewed. That is a conflict the PR page cannot resolve by existing.

Code review has to move to an independent service layer. Not a tab inside the editor. Not a bot comment generated by the same model that wrote the diff. A layer that reads the full codebase, applies verification across security, reliability, code quality, API correctness, ops, and UX, runs those checks in an isolated environment, and submits findings regardless of which forge hosted the code or which model generated it. Hyrax does exactly that: it reads the codebase, finds problems across six domains, runs 13 verification steps in an isolated worktree, and submits a PR. The engineer merges. Hyrax does not care whether the underlying repo lives on GitHub, Origin, or anywhere else.

The what actually changes when a team adopts an AI coding tool post covers the specific patterns that emerge in week one of AI adoption. Generation rate climbs fast. Review time climbs harder. The trend does not reverse as agents become more capable at writing code; it accelerates.

The ownership question Origin hasn't answered#

Three days before Origin shipped, SpaceX completed its $60 billion all-stock acquisition of Anysphere, the company behind Cursor, per an 8-K filed August 14, 2026. Cursor now operates inside a division called SpaceXAI. The vendor asking to hold proprietary source code became, in the span of a long weekend, a unit of a company with its own frontier-model division.

Jason Andersen of Moor Insights & Strategy framed the model-routing question to Tech Times in June: "xAI's models and treatment of guardrails are very different than what Cursor has stood for." That question is sharper now that the deal is closed.

Origin's pricing, security architecture, data-handling terms, retention policy, residency commitments, and subprocessor list were all unpublished at launch, and remained so as of August 18. The changelog says only that Origin reaches "all paid plan users starting today, except enterprise orgs whose admins opt out." Opt-out by default, not opt-in. An organization that has not made an explicit decision about mirroring proprietary code to a new host has had that decision made for it unless an administrator acts.

There is also a separate track record to factor in. In July, researchers at Mindgard disclosed that Cursor would execute a malicious git.exe planted in a Windows project root the moment a user loaded it, with no prompt. Cursor declined to patch it, classifying the issue as out of scope under a shared-responsibility model. No CVE was issued. A vendor that declined to patch a repository-poisoning flaw is now asking to hold repositories.

What to check before Origin enters the toolchain#

The sync architecture makes Origin cheap to evaluate. It is also the property most likely to erode as Cursor's incentives shift toward owning the substrate rather than borrowing it. Ask what egress looks like now, while the mirror is still a mirror.

The default deserves a decision this week, not next quarter. Enterprise administrators can opt out; the question is whether they know to. The data terms deserve a contract, not a product page. And review must not be delegated to the same stack that generates the code. That last point holds regardless of which forge wins.

GitHub's 6h42m outage resolved at 20:22 UTC on August 17. Availability problems have timestamps. The question of who holds your source code, what they may do with it, and who they ultimately answer to does not.

Hyrax is live at hyrax.dev.