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:

• 
  corvus (@CornixCorone) reported 3 minutes ago

  @Gamingtronium had a bug that auto mode in github copilot (which usually uses the cheap but reliable openai models) had trouble with, so i turned on opus 5 and it burned through nearly 50% of my monthly token allowance in one day 😭 lucky that i have 2 weeks of PTO this month. never again

• 
  Pratik Desai (@_Pratik_Desai_) reported 3 minutes ago

  Bad: "description": "searches GitHub" Good: "description": "Searches GitHub Issues by keyword. Returns 10 most recent results with title, status, author and date. Use when asked about bugs or open work in a repo." Same function. Completely different agent behaviour.

• 
  Shadowling McBeard (@Shadowling000) reported 4 minutes ago

  @CyberEagle1989 Wasn't Mythic Mayhem also abandoned? I can see a couple of open issues in github and last patch was 2 years ago.

• 
  Doron Passal (@doronpassal) reported 9 minutes ago

  @htormey Notion and Linear get the meeting because the problem fits on a slide. GitHub falling over at agent scale never gets a slide. The budget still follows whoever was in the room.

• 
  Mahadev Mukherjee (@MahadevMuk52562) reported 10 minutes ago

  Stack underneath all this is Axum on Rust for the API layer, Tokio for async, SQLx so every query is checked against the real schema at compile time, not discovered broken in production, and Postgres holding the actual plot, wall, and ledger data. Deployed on Railway with continuous deployment straight from GitHub, so a fix goes from code to live in a few minutes.@www.growwithmahadev.com

• 
  Iqramul Hussain (@iamiqrram) reported 10 minutes ago

  @cursor_ai @github "Refactor this repo into oblivion." 🤖💥 Meanwhile, GitHub’s official post-mortem: "An internal Kubernetes sidecar failed, triggering a 10x traffic spike from VS Code retry loops." Translation: GitHub didn't get defeated by rival AI. It got taken down by its own IDE refusing to

• 
  Zevryn (@Zevryn0) reported 10 minutes ago

  19 months ago, ai16z was the AI x crypto story. Launched on @Solana in October 2024 with ~$75K raised, it went from a tiny experiment to nearly a $2.5B market cap by January 2025. The pitch was simple: an AI-run venture fund. Then reality caught up. A forced rebrand to @ElizaOS. A 10x token supply expansion in late 2025. A class-action lawsuit alleging the project wasn’t as autonomous as marketed. And on August 4, founder Shaw Walters effectively declared the token dead, with the foundation winding down. The token is now down more than 97% from its peak. But here’s the interesting part: The technology didn’t die. ElizaOS’s GitHub is still active. Development continues. Walters says he’ll keep building the open-source software just without another token attached to it. That’s the lesson I’d carry into the next AI agent cycle: The protocol and the token are two completely different bets. A project can build useful technology while its token goes to zero. Don’t confuse owning the technology with owning the ticker.

• 
  Jeet (@jeetvanii) reported 12 minutes ago

  then next time github is down , they would manually need to fix it

• 
  Jeetendra (@jeetendrayd) reported 12 minutes ago

  @RhysSullivan This was the reason the GitHub server was down, as more code is being pushed due to coding agents

• 
  MJY (@omninomsky) reported 15 minutes ago

  @Tech_girl How would it be able to determine whether the code was generated by AI or people? Realistically what would happen if GitHub would have to spend resources to determine whether the code was generated by a person (sounds expensive). It's very easy to defeat the purpose by being too sensitive (blocking human posts) too lax (allowing too much AI code through) or simply incorrect (blocking humans, allowing AI). It's very hard to get it exactly right. So just from the technical problem alone they probably won't do this. But let's say they did. How does that work out? Well, now people who rely on AI to do their work can no longer use GitHub. They'll probably want to keep using AI so they'll no longer use GitHub. I'm not yet seeing any upside. Less merges per day maybe? Relief for the system?

• 
  Kavi AI Finance (@KaviFinance1) reported 18 minutes ago

  One Hermes trick I wish more people used: Stop putting everything into memory. Use memory for facts. Use Skills for procedures. For example: ❌ Memory: “Deploying my app requires Docker, GitHub Actions, SSH…” ✅ Skill: /deploy-production Then Hermes loads the procedure only when needed. This keeps your persistent context smaller and makes the agent much more consistent. Even better: If Hermes just completed a complicated task that took 5+ steps, tell it: “Save what you just did as a skill.” Now you’ve turned a one-time solution into reusable procedural memory. That’s one of the things I really like about Hermes. You’re not just giving the agent instructions. You’re slowly building its toolbox. And there’s actually recent research backing up why this matters: a new study found that skills often help agents by acting as procedural anchors that stabilize execution, rather than simply injecting more knowledge. Try this one over the weekend: Do something complicated → make it work → save it as a skill → reuse it next time. Your agent gets more useful without your prompt getting bigger. That’s the part of agentic AI I think people are still underestimating. LinkedIn Weekend Hermes trick: don’t confuse memory with skills. This sounds simple, but it completely changed how I think about building a long-running agent. A lot of people try to make their AI “remember everything.” That’s usually not what you want. In Hermes, I like thinking about it this way: Memory = WHAT Skills = HOW For example, your memory might contain: “My production server runs Ubuntu.” That’s a fact. But: “How I deploy my production application” is a procedure. That belongs in a Skill. So instead of stuffing your memory with hundreds of instructions, you can create something like: /deploy-production and let Hermes load the full procedure only when that workflow is actually needed. This has another interesting effect. Your agent starts developing a procedural toolbox. You solve a difficult problem once. You get the workflow working. Then tell Hermes: “Save what you just did as a skill.” Now the next time you encounter a similar problem, you’re not starting from zero. You’re reusing the accumulated procedure. Hermes’ documentation specifically recommends creating a skill for tasks that take multiple steps and will be repeated. Its skills system also uses on-demand loading, so the full procedure doesn’t need to sit in every context window. And this isn’t just a Hermes-specific idea. A recent research paper looking at agent skills found something pretty interesting: Skills appear to help agents primarily by acting as procedural anchors that stabilize execution. In other words: They’re not necessarily making the model “know more.” They’re helping the model execute better. That’s a really important distinction. If you’re building an agent that you plan to run for weeks or months, I’d rather have: small persistent memory + a growing library of good skills than: one massive memory file containing everything I’ve ever told the agent. My weekend experiment for anyone running Hermes: Pick one annoying workflow you’ve already solved. Make it reliable. Turn it into a Skill. Run it again next week. Then improve the Skill when you discover an edge case. Do that 20–30 times and you aren’t just using an AI assistant anymore. You’re building a procedural operating system around it.

• 
  Senb0n22a (@senb0n22a) reported 20 minutes ago

  @0xIlyy if there's obfuscation they don't want you reading the code, could have trade secrets or things of that nature. no issue with decompiling code that's just plain machine readable like exes compiled from github, but you aren't meant to know if it's an encrypted program.

• 
  Sayem (@Sayem314) reported 23 minutes ago

  @wiretransfer @theo srt and vtt, you can request other formats in github issue. should not be hard to implement.

• 
  Julian Goldie SEO (@JulianGoldieSEO) reported 23 minutes ago

  Cursor Origin launched with some pretty wild timing. GitHub went down for roughly 6.5 hours around the same period. And error rates reportedly climbed hard across parts of the platform. That made Origin’s value proposition obvious: → Your code can live in Cursor → GitHub can remain synced → AI agents can work inside the repo → Pull requests stay visible → A second environment gives you another option Origin is still early. But redundancy suddenly sounds less boring when your main code host goes dark. Save this video, you’ll remember why a second code home can matter. Want the SOP? DM me. 💬

• 
  godgiven (@bygodgiven) reported 28 minutes ago

  The Hugging Face issue asking you to upload your weights was not written by a person. Niels Rogge automated his own job at Hugging Face. His agent opens hundreds of GitHub issues a night and never says it is an agent. His reason, from the talk: if people know it is a bot, they close the issue. Two negative comments across thousands of issues. He now watches people reply to his agent with their own agent. The builder beat is not the outreach. Version one automated the writing and handed him the inbox, so the bottleneck just moved to replies. Version two is one CLI, one skill, one sandbox, running on somebody else's containers. Cursor's David Gomes, same conference: 12,000 lines of feature code replaced by a 200 line skill. Open your issue tab. Check who actually wrote them. Who owns the loop underneath is in the article below.