SShortSingh.
Back to feed

OpenAI, Anthropic and Cursor's SpaceX Build Their Own Code Hosting as GitHub Struggles

0
·2 views

Several AI companies are building their own code hosting systems in 2026 rather than relying on GitHub. Cursor, now owned by SpaceX, has launched Origin, an agent-first Git host built into its editor, while OpenAI is developing Pando and acquiring Pierre Computer to bring Git hosting into Codex. Anthropic has replaced Git for its monorepo with an in-house system called Scion. The shift is driven by the heavy volume of code generated by AI agents, which has contributed to frequent GitHub outages, including 257 recorded incidents from May 2025 to April 2026. GitHub remains widely used and is expanding its capacity as it migrates to Azure.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Bolt CMS Flaw CVE-2026-77380 Allows SSRF via Unchecked Redirects

A server-side request forgery vulnerability, tracked as CVE-2026-77380, affects Bolt CMS versions prior to 6.1.5. The issue lies in the asynchronous URL upload and media embed endpoints, where the initial user-supplied URL is validated against internal network ranges but HTTP redirects and meta refresh directives are followed without re-validation. An authenticated user with editor privileges could exploit this through an open redirector or DNS rebinding to reach internal network resources and cloud metadata services. The vulnerability carries a CVSS score of 5.2 and was published on 2026-11-20. Administrators are advised to upgrade bolt/core to version 6.1.5 or later, audit editor-level accounts, and restrict outbound connections from the web server to private address ranges.

0
ProgrammingDEV Community ·

RipSeek's MCP server earns top quality grade on Glama and joins official registry

RipSeek, a Collector Crypt pack product, runs a remote, read-only MCP server that lets AI apps like Claude, ChatGPT and Cursor query pack odds, valuable cards and recent pulls. Built as a stateless Streamable HTTP endpoint on a Cloudflare Worker, it exposes six single-purpose tools and cannot buy or sell anything on a user's behalf. After a review by Glama, the server was rewritten with clearer tool descriptions, a new single-card lookup and expanded filters, and it earned an A grade (3.9 out of 5) on tool quality. The team then published its server.json to the official MCP Registry, from which directories such as Glama pick listings up, and it appeared on Glama within a day.

0
ProgrammingDEV Community ·

Solana web dApps wrapped as Seeker apps face Google sign-in and wallet hurdles

A team behind six Solana dApp Store apps describes converting web dApps into Android apps for the Seeker phone using Solana Mobile's free WebView shell tool. The tool builds an APK from a web manifest and is now preferred over the older Trusted Web Activity route, though wallet adapter packages need updating to 0.5.1 or later. The main problem they found was Google sign-in: links outside the app's host open in the system browser, leaving the app stuck on a loading screen after authentication completes in Chrome. Their suggested fixes include returning to the app via deep links or App Links, making the Seed Vault wallet the primary sign-in method, and reviewing every window or pop-up flow. They also recommend testing wallet connection, transaction signing, declined signatures, and Android back-button behaviour on a real device.

0
ProgrammingDEV Community ·

Kabloom: A Daily Trivia Game Where Each Correct Answer Can Only Be Used Once

A DEV Community article examines Kabloom, a free daily browser trivia game with seven clue checkpoints on a pixel-art volcano. Players type names that satisfy each clue and climb when the answer is accepted. Once a name is accepted, it cannot be reused later in the same run, which turns early answer choices into a strategic trade-off. Each clue has a brief reading pause followed by a 25-second answering window, and wrong guesses do not end the climb immediately. The article argues this single-use rule links the seven questions together and makes knowledge function as a limited resource.