Cursor launched a GitHub competitor the day after GitHub fell over for eight hours — here's whether Origin is worth your repos

What actually shipped

Origin is a code host. That plain description matters, because the interesting claim is not that Cursor built something novel in kind — it’s that a company known for an AI code editor decided the forge itself was the thing worth rebuilding.

The product page carries a single line of positioning — “A git forge for the agentic era” — and one line of availability: early beta, now available on all paid plans. Feature-wise it covers the recognisable surface of a forge: repositories, browsing and editing code, pull requests, and collaborative work on codebases. Cursor says “agent native” capabilities and a broader “app ecosystem” are coming, but has published no detail on either, so neither belongs in a buying decision today.

The interoperability story is the part worth reading twice. Origin is designed to sit beside GitHub rather than replace it: “Your GitHub repos can sit alongside the ones Cursor hosts,” Cursor writes. “Connect GitHub to Cursor, pick your org, and you’ll see the repos you can sync. Select one and Cursor pulls it in.” For a challenger in source-code hosting, that is the only sane opening move — the switching cost of a forge is not the repository, it’s the twenty integrations hanging off it.

The outage that wasn’t quite the same day

The framing in most coverage is that GitHub fell over on launch day. The dates don’t quite support it, and the real numbers are more damning than the paraphrase.

Per GitHub’s own incident history, the major incident ran on 17 August 2026 from 13:28 to 21:15 UTC — 7 hours 47 minutes. At peak, web and API error rates hit roughly 20%, while archive and raw-content downloads reached approximately 50%. Issues, Pull Requests, APIs, Actions and Copilot were all affected, along with SAML/OIDC authentication, SCIM and Team Sync.

The root cause is worth quoting in outline because it is a very modern failure: network saturation on load balancers in GitHub’s Central US datacenter, triggered when an Istio sidecar pod hit its concurrency limits and failed to autoscale because of a policy watching the host service but not sidecar limits. That cascaded until four HAProxy nodes exhausted their flow limits, degrading the gateway auth path. Then it got worse — “optimistic retry logic which overloaded internal load balancers,” plus a latent retry bug in VS Code that amplified traffic by approximately 10x, delaying recovery of the Copilot Token Service until 21:02 UTC.

So: not launch day, but the day before — and a genuine multi-hour, multi-service failure with a retry storm on top. Cursor’s post landed on the 18th, into an audience that had spent the previous afternoon unable to merge anything. TechCrunch also cites a LeadDev analysis counting 257 GitHub outages over the past year; we haven’t independently verified that count, so treat it as reported rather than established.

None of this means Cursor timed anything. A 27-minute engineering write-up describing a bespoke storage system is not a week’s work. The outage sharpened an existing pitch.

Continuity: the actual engineering claim

The substance of Origin is a storage system called Continuity, described in detail by its author, Vicent Martí, who previously worked on version-control infrastructure at GitHub. The write-up spends most of its length on why the obvious approaches fail, which is unusually honest for a launch post.

The core problem: Git stores objects in packfiles, and the protocol requires packfiles on the wire regardless of how you store data server-side. Distributing the filesystem under Git — NFS, GFS, block-level replication with DRBD — was tried at GitHub and, per Martí, repeatedly hit a wall, because Git makes assumptions about local filesystem semantics that network filesystems violate. Distributing Git at the object level into a key-value store fails differently: a repository is a DAG, and you can’t know the next pointer until you’ve fetched the previous one, so every operation becomes a chain of round trips.

Continuity’s answer:

• A write-ahead log in S3-compatible object storage is the source of truth. A push is stored as a WAL entry and never acknowledged until fully persisted. It becomes visible only once the reference transaction is prepared and a pointer is recorded in the WAL index — which forces pushes to be linearizable.
• Repositories on disk are a warm cache, not state. If a repo is missing on a host, it is materialized from the WAL. Location is decided by rendezvous hashing over healthy nodes — no routing tables, and, in Martí’s aside, no relational database to operate.
• No consensus, no elections. Any server can act as primary for any repository; safety comes from an atomic compare-and-swap on S3. In the healthy case one node is consistently chosen for efficiency; in a deploy, failover or network blip, it simply doesn’t matter which.
• Optimistic replication over UDP gossip, with replicas catching up directly from S3 — leaning on object-store scalability rather than a replication topology.
• The local copy remains a normal Git repository on NVMe, deliberately, so upstream Git and its optimizations keep working.

The stated design goal — “always correct when degraded, and always fast when healthy” — is the right target for a forge, and the architecture is a coherent way to aim at it. It is also, at this stage, entirely Cursor’s own account. There is no independent uptime record, no third-party audit, and no public status-page history to weigh against GitHub’s 257 reported incidents. Architecture is a reason to take something seriously; it is not evidence of reliability.

Why this matters

The agent-volume argument is the real thesis, and it’s sound. Cursor’s framing is that agents changed the load profile of version control: “More code, more PRs, more CI runs.” That matches what teams running Claude Code, Codex or Cursor’s own cloud agents actually report — machine-generated branches and pull requests arriving at a rate human review workflows were never sized for. If you believe agentic coding keeps scaling, the forge genuinely does become a bottleneck, and “version control is possibly the hardest thing to change overnight” is a fair warning. That argument would hold even if GitHub had a perfect year.

Vertical integration is now the competitive move in coding tools. Cursor already ships an editor, a model router, cloud agents, and — since Grok 4.6 — deep model ties through its new owner. Adding the forge means one vendor holding your editor, your agents, your models and your repositories. That is a real efficiency story and a real concentration risk, and the Cursor vs Claude Code and GitHub Copilot vs Cursor comparisons now carry a question they didn’t a week ago: how much of your toolchain do you want under one roof? Note that GitHub Copilot’s owner has the same integrated position, from the opposite direction. Two days after Origin’s launch the opposite bet arrived: Warp opened early access to Factories, agent pipelines whose whole premise is that each stage should be free to run a different model and harness. Whether the durable position is owning the stack or abstracting over it is now a live disagreement with products on both sides.

Ownership changed four days before launch. Cursor confirmed on 14 August 2026 that it is now part of SpaceX. Whatever you make of that — and it cuts both ways, bringing capital and infrastructure alongside a new set of strategic priorities — it is a material fact when deciding who hosts your source code. A forge is a decade-long commitment; the entity behind this one is four days old in its current form. The appetite has not stopped there: Bloomberg reported on 19 August 2026 that SpaceX approached Cognition, maker of Devin, about an acquisition — a report Cognition CEO Scott Wu publicly denied, stating the company is not for sale and has not been in acquisition talks, with the two now described as discussing compute rather than a sale. Treat that as a rumour with a named denial attached, not a transaction; treat the pattern it sits inside as real.

GitHub’s moat is the ecosystem, not the storage layer. Even granting Continuity everything it claims, GitHub’s ~180 million developers come with CI providers, security scanning, dependency tooling, package registries, compliance attestations and institutional muscle memory. Origin’s sync-alongside design is an acknowledgment of exactly this. Buyers evaluating the coding-tools landscape should weigh integration breadth at least as heavily as push latency.

The verdict

Origin is a more serious product than the “GitHub had a bad day, Cursor pounced” framing suggests. The engineering is specific, the failure modes it targets are the ones that actually bite forges at scale, and the person who built it has direct experience of the systems he’s critiquing. The interoperability-first design shows a company that understands it is the challenger.

It is also an early beta, from a company four days into new ownership, whose promised differentiator — agent-native workflows — has not been described in any detail. The reliability case rests entirely on self-reported architecture. Cursor filled in part of that blank the following day: its 19 August cloud-agent release added event-driven Subscriptions, persistent /goal objectives and isolated subagent VMs — the agent-side half of the same bet, and the reason Origin’s “more code, more PRs, more CI” argument is not hypothetical.

Recommendation: if you’re on a paid Cursor plan, put a non-critical repository on Origin and exercise it — sync from GitHub, run pull requests through it, see how it behaves under your actual agent volume. That costs you nothing and tells you more than any launch post. Do not move your organization’s source of truth, and do not plan a migration until the agent-native features exist as documented behaviour rather than a promise. Keep watching two things: whether Cursor publishes a public status page with a real incident history, and whether the GitHub reliability trend that made this launch land continues into the autumn. If both go Cursor’s way, this stops being a curiosity and becomes a genuine procurement question.

Questions or corrections? Email Pick Right. Want the full list? See all news.