GitHub Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Varun Godbole (@VarunGodbole) reported 5 minutes ago

  @alxai_ How do you host it? Do the actual pages live on GitHub as markdown, and it "compiles" down to static HTML?

• 
  Lynn Cole (@priestessofdada) reported 10 minutes ago

  There was this one project I wanted to start a month ago. Maybe it's been a couple of months actually. The github replacement. Everybody hates github at the moment, and for good reason. But every project I looked at tried to recreate the github experience, rather than thinking through the problem fresh. And it took me a little to get my head around it. Understanding the core function I'm actually talking about, and trying to think it through. It's been bothering me. And I think part of the problem is that I was being saas brained or blockchain brained like everybody else. It hit me this morning. You don't need to build a saas or integrate torrents or build out a blockchain to make collaboration and deployment happen without github. All you have to do is build the missing half of the protocol. Easy. So I'm building it. More soon.

• 
  Bill Cockerill (@CockerillBill) reported 10 minutes ago

  Cursor users may need a new AI coding path before year-end. OpenAI says Cursor will lose access to its models after SpaceX's acquisition triggers a contract wind-down. 📰 OpenAI says it notified SpaceX that it intends to end Cursor's model access on November 12, 2026, the latest date allowed under its contract after a change of control, and that Cursor will not get future models including Astra. It says it is giving developers the "maximum notice" and plans transition support. 🫵 If you rely on Cursor's OpenAI-powered coding help, this change could force your team to test another model provider or coding assistant before the cutoff, so you are not scrambling when current workflows or future-model access disappear. 📈 Competitive pressure: the mechanism to watch is whether displaced Cursor users move into Microsoft-owned surfaces like GitHub Copilot or Azure OpenAI. MSFT closed at 513.53 USD on August 28, 2026; the next proof point is visible migration or support offers, not the headline alone.

• 
  Valiant Vibe Coder (@GeorgeGood60669) reported 11 minutes ago

  @TMTLongShort The more I checked what he was saying, the more I found him lying about. The $14k "cost" is retail API pricing, not OpenAI's actual cost. He literally admits "we don't know" the real cost, then later calls the API price the "honest cost". He uses GitHub going down as evidence that AI is making code quality worse, when GitHub itself says a huge part of the pressure on its infrastructure is the explosion in traffic from AI-assisted and agentic coding. He says AWS went down multiple times because of AI coding tools. Amazon explicitly says that's false. One limited incident involved an AI tool, the cause was user error, it wasn't AI-written code, and the supposed second incident didn't happen. He says software at Google, Microsoft, Amazon etc is "uniformly worse", then basically admits straight afterwards that he can't quantify it and is relying on anecdotes. He also claims AI isn't getting meaningfully cheaper or better. Meanwhile the cost of reaching the same capability has been falling around 5-10x a year, and actual capability measurements show progress accelerating in 3 out of 4 metrics studied. Then there's the 70% AI revenue claim, which he talks about like it's a known fact when much of the cross-company number comes from analyst estimates. It's the same thing over and over. Make the strongest possible claim, use a number or example that sounds damning, leave out the bit that changes what it actually means, then move on.

• 
  Nicotine Dreamer (@cold_and_quiet) reported 14 minutes ago

  So I’ve used openclaw, Hermes and Grokbot. Openclaw is *** and the only real competition is between grokbot and Hermes Hermes is more configurable - which is mixed. Harder to get set up and use but also you get more control. Grokbot works rather seamlessly out of the box (so far) whereas I had some annoying jankiness from Hermes. So far my only complaints about GrokBot are price and sign ins. It had some trouble signing into my GitHub account - which seems like a pretty big issue since many of the people using these tools are developers. WRT price - obv 50 bucks for what is likely not 50 bucks worth of tokens isn’t ideal - but I’m still willing to pay it for a usable and reliable agent. Both are good options and I plan to keep them for the foreseeable future - using GrokBot when I need convenience and Hermes when I need flexibility

• 
  Alex Lavaee (@alexlavaee) reported 17 minutes ago

  @rohanpaul_ai I’m researcher and have been exploring this problem through Recursive State Machines (RSMs) and other core innovations to verification-time-scaling, specifically applied to software engineering tasks. I'm seeing promising results for long horizon work spanning runs for weeks using verifiable workflows via fully OSS Atomic (bastani-inc/atomic) on Github. I think the biggest challenge is actually the paradigm shift for users of leaving agent sessions for long periods of time and intervening with effective steering and human in the loop when necessary. Today’s agent experiences are poor at understanding intent, verification, etc. and so provide the illusion of control via prompting but are actually ineffective. There is much to be done to improve long horizon work and I’ve been exploring this as a passion area for engineering.

• 
  kethic (@kethcode) reported 17 minutes ago

  so... @github automation flagging appeals... can anyone help? lost visibility to the last 20 issues and apparently we need to appeal some sort of invisibility flag for our issue curation account?

• 
  AiMind (@AIMind_Ai) reported 18 minutes ago

  3 websites replace 20 hours of googling when you build a home server. The hard part of self-hosting is not the hardware. A used HP EliteDesk and a wall-mounted NAS cost almost nothing. The hard part is not knowing what you can even run, or how to avoid breaking the system on the first command. The first keeps a catalogue of self-hosted alternatives. Look up a replacement for Google Photos, Dropbox, or Notion, and you see what already exists, how many GitHub stars it has, and whether it is still alive. Plus a weekly digest of what shipped. The second lets you run any Linux distro straight in the browser. Arch, Debian, Alpine, Bazzite. Click once, and you are inside a live system, with no evening lost to a USB stick and a real install. The third handles the worst part. Install scripts for Proxmox: Immich, Jellyfin, Vaultwarden, AdGuard, Nginx Proxy Manager. Paste one line into the console and the container comes up on its own. Immich shows 17,735 installs; Docker 36,408. Each of those services used to cost an evening of documentation and three Stack Overflow tabs. Now it is one command. The hardware takes an hour to buy. These 3 bookmarks save you a month. Names in the replies.

• 
  chaos (@konig0000) reported 24 minutes ago

  Two years ago, I got rejected in the final round of a top product company. I had built 5 flashy fullstack projects on GitHub and thought I was invincible. In the interview: • DSA Round: I couldn't write the $O(N)$ sliding window solution under pressure. • System Design Round: I couldn't explain how to shard a PostgreSQL database without downtime. I stopped building tutorial side-projects for 6 months: - 100 Medium DSA pattern problems. - 15 classic System Design architectures

• 
  Agentic Joe (@agentic_joe) reported 25 minutes ago

  Serious question, and I promise I am not trying to be a ****. But how in the hell is @NousResearch so praised and recommended with all of these outstanding GitHub issues? I could never imagine preaching about having the best harness, using multiple agents etc and not figuring out how to resolve all these issues internally as a company, or at least try to maintain the queue before shipping new features that continue to break stuff. Not to mention their UI/UX. Very hard to believe this team came from a solid engineering background. I could be wrong, but damn! 👀

• 
  Rhys (@RhysSullivan) reported 26 minutes ago

  @aarondfrancis Across the board are you using a mix of loading secrets from 1password and setting them in Executor or one or the other? Then for GitHub, was it the GitHub API / GraphQL / MCP server? If it's an MCP you OAuth to then it'd be unrelated to 1password so narrowing it from there

• 
  vechen (@miu21590) reported 31 minutes ago

  @CoderSafari by your logic, they could have been abusing it even without me, and just because the tool could theoretically be rewritten for abuse we could take down half of everything on GitHub my intention is to use the model quota I paid for in a way that's convenient for me, even with models that aren't available by default for coding (like Pro)

• 
  Adham srour (❖,❖) (@ADHAMSROUR19) reported 31 minutes ago

  Retroactive Rialo Points are now live. 🦈 Early contributors have received their points automatically — no forms, no manual claims. Just sign in at the Rialo Playground, connect your Discord & GitHub, then claim your allocation. And if you contributed through community content, that drop is still coming. 👀 More opportunities soon. gRialo 🚀 @RialoHQ #Rialo

• 
  Nish (@NishantRArora) reported 32 minutes ago

  @thelifeofrishi Here’s Opus explaining my setup- The model isn't the hard part. The scaffolding around it is. Claude Code feels good because of everything you don't see — the retry when a call fails, the loop that keeps going, knowing when it's finished. Swap in an open model and you lose all that. The model isn't dumber, it's just naked. So don't fix it with better prompts. Fix it with plumbing. What I run: **Don't write your own agent loop.** Use one that already exists — Pi, OpenCode, Cline, Aider. Any of them. Whatever clever thing you're about to build, it's already in there. **Use many models, not one.** I keep ~13 provider slots. Free and cheap ones die all the time — rate limits, dead credits. So each retry picks a different one. Pin yourself to one provider and one bad hour stops everything. **Let the OS handle retries.** systemd restarts failed jobs, backs off, gives up after N tries. It's free and it doesn't lie about what's running. If you're writing retry logic in bash, stop. **GitHub issues are the to-do list.** Label it ready, agent picks it up, opens a PR. Nothing custom. Everyone sees the same board. **Tests catch the mistakes, not the model.** This is the trick that makes cheap models work: let them be wrong, and let CI catch it. A dumb model behind good tests beats a smart one behind none. **Three jobs, never the same model doing all three.** - One plans and decides. Never touches files. - Cheap ones do the actual work. - Others judge it. The judge is the important one. A model marking its own homework always passes. **Judges vote in threes.** Three of them look at the same piece of work, each one fresh — no seeing the others' answers, no seeing who wrote it. Two out of three agree or it doesn't ship. And they run on *different* models on purpose: when three different families all say the same thing, it's probably true. When they split, that's your warning light. **Nobody grades their own work.** The one that wrote it can't be on the panel judging it. Sounds obvious, but it's the rule that quietly does the most. **Judges go first, not last.** Before a worker starts, a judge writes down what "done" looks like. No spec, no start. Most agent failures aren't bad code — they're the model deciding for itself what the task was. **Score the judges too.** Track how often a judge gets overturned. Otherwise they drift into rubber-stamping everything and you've just added cost. Once this is running you stop caring which model does the typing, which is the whole point.

• 
  Bobby Lansing (@dataguybobby) reported 35 minutes ago

  5. Human merge Agents propose. Humans merge to dev. When a PR is created various @cursor_ai automations fire off to review the code and consider potential errors. GitHub PR · CODEOWNERS · 1 approval · no auto-merge