Cursor Origin: SpaceX Attacks GitHub With a Code Platform at a Weak Moment

Set Trending Topics as a preferred source on Google.

Cursor now hosts code as well. The startup behind the most popular AI editor launched its Origin product as an early beta this past Monday and is rolling it out gradually to all paying users. That puts a company into the code hosting market that so far mainly depended on others running it, GitHub above all. The timing, the data question and the new owner make the launch more than a feature update.

What Origin can do

Origin shows up inside Cursor as a separate tab called Codebase. From there, users can create new repositories, clone them via CLI or push local projects. Every repo comes with pull requests including timeline, commits, checks and diffs, plus comments and a merge function. The name of the codebase becomes part of the URL, following the pattern cursor.com/codebase/acme-corp.

GitHub synchronisation is central to the product. Anyone who wants to can connect their GitHub organisation and mirror individual repos to Origin. Those copies update in real time and can be browsed and downloaded, but pushes still go to GitHub. For synced projects, GitHub remains the source of truth, according to Cursor. Pull request comments sync in both directions, so a review can be handled and merged in Cursor while the rest of the team stays on GitHub.

On top of that comes an emerging app ecosystem: Vercel delivers preview deployments per pull request, Depot and Buildkite handle CI and run existing GitHub Actions workflows. The genuinely interesting functions are still missing, though. Cursor merely announces agent-native features, while what shipped so far is essentially a mirror with an interface. Enterprise organisations can have their admins switch Origin off, for all other paying accounts it is enabled by default. Details on scope and setup are in the Cursor changelog.

Why now of all times

The launch landed on a day when GitHub went down worldwide. According to The Register, error rates for downloading repository content reached around 50 percent, with issues, the API and Copilot also affected. Microsoft confirmed the disruption. Cursor has not commented on whether the timing was coincidence or calculation, in the developer community it was commented on extensively.

The outage does not stand alone. GitHub has been struggling with reliability problems for months, which the company itself partly attributes to the load created by AI-assisted and agentic coding. Software engineering chief Jakub Oleksy announced structural changes in early summer to permanently remove failure modes. In parallel, GitHub Copilot is losing market share: analyses of the Stack Overflow Developer Survey cited by industry media see its share falling from roughly two thirds to about half within a year, while Cursor and Claude Code gain ground. Cursor itself says it has pushed annual revenue to two billion dollars.

Attacking GitHub is still no small undertaking. The network effect of issues, Actions, marketplace, open source history and identities cannot be rebuilt in a beta. Cursor is therefore betting on coexistence rather than migration for now.

The connection to the models

Origin gets interesting where it touches the business with AI models. Cursor develops its own coding models, and the current Composer was built on the open Chinese base model Kimi K2.5 according to the company’s own technical report, then trained with large-scale reinforcement learning, specifically in realistic Cursor sessions using the same toolset the deployed model uses. For evaluation, the company relies on CursorBench, a benchmark drawn from real coding sessions of its own engineering teams. The fact that a self-developed model rests on a Chinese foundation has been a talking point recently.

This is exactly where the strategic value of code hosting lies. Anyone who knows not just the editor but also sees repos, pull requests, reviews, test runs and CI results can build training and evaluation environments that come far closer to actual working reality than chat logs alone. A pull request delivers task, solution, discussion and a success signal in one package, in other words the raw material for agentic training. Cursor already operates an internal platform with hundreds of thousands of sandboxed environments for that purpose.

How Cursor handles code stored in Origin specifically remains open, however. The general data policy distinguishes between privacy mode enabled, where according to the company neither Cursor nor the model providers train on the data, and privacy mode disabled, where codebase data, prompts and code snippets may be used to improve AI features and train models. Dedicated terms for Origin, covering retention periods, data residency, subprocessors or export tools, have not been published by Cursor so far. The Next Web and Tech Times point out that the product is switched on by default at the same time, and advise using Origin in mirror mode only for now.

New owner: SpaceX

The launch falls into the first days under a new flag. SpaceX closed its 60 billion dollar all-stock acquisition of Cursor parent Anysphere last week, after the group had secured the right to do so in spring. Cursor thereby lands in the AI unit SpaceXAI, into which xAI had already been folded. Cursor argues primarily with compute and points to access to what it describes as the largest GPU fleet in the world.

The interlocking is already visible. The Grok 4.5 model was trained jointly with Cursor according to SpaceXAI and was initially unavailable in the EU. With the most recent joint product, Grok Bot, the Cursor brand already receded into the background.

What emerges is a constellation that did not exist in this form before: one group controls the editor, the models, the data centres and now also the platform where the code sits. For users that means short paths between tool and model, but also a concentration in which questions about data use, lock-in and governance are harder to answer than with separate vendors. As long as Cursor does not publish data terms for Origin, GitHub remains the authoritative place where the code sits for synced projects anyway.