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:

• 
  Nathan Brake (@natebrake) reported 2 minutes ago

  @victorsavkin It's like the sound guy at a concert, when they're doing their job, no one notices them. Agree, github errors have dropped and unfortunately for them it just means we stop complaining about them, I don't see much chatter about 'omg amazing, we love github now' :(

• 
  regent0x (@regent0x_) reported 2 minutes ago

  Anthropic CEO Dario Amodei on how they use claude internally: claude now writes 90% of their code and handles 77% of real github PRs, up from 5% a year ago a bug broke their whole training cluster, engineers chased it for days, they told claude "just poke around" and it found it in one prompt but the claude on your machine can't do this, out of the box it only reads files and runs terminal commands, no PRs, no database, no slack it's a genius locked in a room with no windows MCP servers are the windows, 4-6 of them in one afternoon turn claude from "writes code when asked" into "reads the failing PR, queries the db, posts the fix in slack" in one prompt most people bolt on 20 broken servers and wonder why claude got dumber the guide below has the exact handful that matter

• 
  Sudo su (@sudoingX) reported 3 minutes ago

  now i'm running the harness fight i've wanted to run for a month. hermes agent vs openclaw, same model, same tasks, both pointed at a 3.9gb bonsai on a single 3090. lean vs bloated, head to head, and i post whatever happens. disclosure first, i contribute to hermes, so i'm not pretending i'm neutral. what i can do is make it a fair fight. upstream versions of both, same local endpoint, no fork tricks, and let the receipts talk. here's what's actually in each box. > hermes agent. a one-file agent loop, around 25 direct dependencies. it parses and repairs tool calls off many model families natively, which is the whole reason it reads what a local model actually emits. built for open weights from day one. #1 on openrouter by daily usage. > openclaw. thousands of typescript files, roughly double the dependencies. for years it leaned on the server to parse tool calls and only shipped its own repair recently, for a single format. way more github stars than hermes. built for the big api models, and it shows. now the honest part. i uninstalled openclaw months ago. my experience was that it was built for someone else's models, it choked on local, and the bloat made it slow just to start. but that was months ago and these things move fast. maybe they fixed the local story. maybe the parsers are there now. i'm not going to assume, i'm going to run it. that's the whole test. can either harness hold a tool-calling loop on a 3.9gb model without falling apart. one early tester says bonsai breaks on iteration, another says it tops agentic benchmarks. that split might not be the model at all. it might be the harness. this is what finds out. results coming. i'm not calling it early.

• 
  riya (@riya_mishra007) reported 4 minutes ago

  10 years ago you needed a team. Today you need a laptop to build a SaaS at $0 to earn more than +$10M. Claude for coding. Supabase for backend. Vercel for deploying. Namecheap for domain. Stripe for payments. GitHub for version control. Resend for emails. Clerk for auth. Cloudflare for DNS. PostHog for analytics. Sentry for error tracking. Upstash for Redis. Pinecone for vector DB. It's not that difficult broooo.... You can literally ship a startup sitting home.

• 
  bizz | AI software solutions (@bizzaidev) reported 9 minutes ago

  2/ The tool: Libretto PR agents — free TypeScript. Add a try/catch line and it pulls the failure session (via CDP), inspects the page with injected Playwright/JS, and opens a GitHub PR proposing a fix. No full runtime migration required.

• 
  Rokko (@0xRokko) reported 11 minutes ago

  Shunyu Yao - creator of ReAct:"A language model is not very good at self-evaluation yet." Read that again. The guy who invented the agent loop is telling you the model cannot grade its own work. Yet everyone is out here letting it mark its own homework and shipping the garbage it gives an A. Here is what actually happens when you stop trusting the model and build the check yourself. ReAct with a HANDFUL of examples hit 40% and beat an RL agent trained on 100,000 samples. On real GitHub issues, plain models solved 2%. Bolt a reason-act-observe loop on top and it jumped past 10%.Same model. 5x the result. The difference was never the prompt. It was the verifier. Your taste, written down strict enough that a machine enforces it, is the whole game now.

• 
  Max ¯\_(ツ)_/¯ (@MaximumADHD) reported 14 minutes ago

  This fixed itself after a few hours, I think it was a GitHub outage.

• 
  Roposh (@Roposh82) reported 15 minutes ago

  𝗜𝗳 𝘆𝗼𝘂'𝗿𝗲 𝘀𝘁𝗶𝗹𝗹 𝗵𝗲𝗿𝗲 𝗶𝗻 𝘁𝗵𝗶𝘀 𝗯𝗲𝗮𝗿 𝗺𝗮𝗿𝗸𝗲𝘁... 𝘆𝗼𝘂'𝘃𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝘄𝗼𝗻. Not because you've made money... But because you refused to quit. Most people won't experience the next bull run. Not because it won't happen... But because they gave up before it arrived. They sold. They lost hope. They called crypto a scam. They walked away. Almost every friend I know has either left the market or spends every day complaining that their portfolio is down 90%. 𝗜'𝗺 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲 𝗼𝗽𝗽𝗼𝘀𝗶𝘁𝗲. I've never felt more prepared. I've never had more conviction. I've never been more excited for what's coming. The next 𝗼𝗻𝗲 𝘁𝗼 𝘁𝘄𝗼 𝘆𝗲𝗮𝗿𝘀 won't just be another cycle. For some of us... It will be the moment that changes our lives forever. People will celebrate the gains. They'll never see the price we paid to earn them. 𝗧𝗵𝗲 𝟯𝗮𝗺 𝗻𝗶𝗴𝗵𝘁𝘀. The endless research. The whitepapers. The GitHub commits. The charts. The countless posts that nobody cared about. The doubt. The ridicule. The feeling of believing when almost nobody else did. My highest convictions are 𝗤𝗨𝗕𝗜𝗖, 𝗧𝗔𝗢, and 𝗞𝗔𝗦. Maybe I'll be wrong. But I'd rather fail following my convictions than spend the rest of my life wondering "what if?" 𝗧𝗵𝗶𝘀 𝗶𝘀𝗻'𝘁 𝗷𝘂𝘀𝘁 𝗮𝗯𝗼𝘂𝘁 𝗺𝗼𝗻𝗲𝘆. It's about proving that patience has value. That conviction matters. That staying when everyone else leaves is sometimes the only difference between an ordinary life and an extraordinary one. One day they'll ask... "𝗛𝗼𝘄 𝗱𝗶𝗱 𝘆𝗼𝘂 𝗸𝗻𝗼𝘄?" And we'll smile. Because we didn't. 𝗪𝗲 𝗷𝘂𝘀𝘁 𝗵𝗮𝗱 𝘁𝗵𝗲 𝗰𝗼𝘂𝗿𝗮𝗴𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆. $QUBIC $TAO $KAS

• 
  OIiver (@posedscaredcity) reported 19 minutes ago

  @my_knn_totoro @KSimback i actually run gstack across my company and can answer this too ( i was just seekign outsider opinion) pros: - works in practice like magic now for us - the agents are continuously learning. the default output before vs after is like a 3 generation model difference on the same model. gpt 5.5 with it was comparable to fable without it. fable with it is insane. - much easier to prompt - no need to transfer much context - new hires and anyone can get any and all questions out of their wheelhouse answered as needed - tracks decision etymology in a way that was missing cons: 1. its quite broken: many days of agent time spent to get and keep it working. dreaming has broken so many times. 2. authentication wasn't developed or wasn't developed well and setting up new hires or new agent systems to hook in with correct attribution is a ***** (with how i set it up at least) 3. once installed agents do not use it and do not use it well. we needed a good agents.md file telling it to look for task preferences before starting, and to fill out the empty search queries from the start when wrapping up and meta preferences within gbrain itself. 4. it slows down the agents since they have more to traverse 5. ingestion was broken out of the box and integrations sucked. we hooked in and heavily modified composio so i could ingest a lot of events 6. connecting a github account will ingest all events from all open source repos you've ever touched. cleaning that up was a ***** 7. federating access is really hard as a result haven't bothered but isn't scalable.

• 
  Denis Volkhonskiy (@den_volkhonskiy) reported 20 minutes ago

  the simple loop that will turn your claude code and codex subscription into a team of engineers 1. add codex code review on github to your repo 2. use claude code for development 3. ask claude code: "create PR, babysit it, check every 5 minutes for comments from codex. If there are comments, validate and fix them. when you see thumbs up reaction on the PR body, finish the loop and merge it"

• 
  Filip (@fristovic_) reported 20 minutes ago

  GitHub is down. AWS sending astronomical bills. Cloudflare bugging out. The end is near.

• 
  Tolys ✨ (@0x0001337) reported 23 minutes ago

  @MetaMask AI checker just ruined our Wallet Connection. Waiting for resolution in github issue

• 
  Hugo Bowne-Anderson (@hugobowne) reported 23 minutes ago

  “You still use pull requests? I wouldn’t even do that anymore. Just push it straight to trunk, have your agent summarize it.” That’s @gregce10, co-founder and CPO of SpecStory. He previously worked at GitHub, Dropbox and Google, and was CPO at Pluralsight. And he kept going: - PRs are the limiting gate when agents produce more code than humans can review. - The model should never decide when its own work is finished. Put the deterministic checks somewhere it cannot access. - *** is probably here to stay. Whether GitHub remains the platform, “we’ll see.” @HanchungLee came at the same problem from the evaluation side. Han is Director of Machine Learning at Moody’s and works on SkillsBench, evaluating skills across combinations of models and agent harnesses. - An agent is the model plus its harness. You need to evaluate the complete system. - A green check proves nothing if the agent found a way to game the task. - Your agent could delete the failing test and declare success. Both are figuring out how to turn masses of agent-generated slop into signal. Greg mined 516 saved agent sessions to recover the decisions and intent behind the work, identify recurring practices, and forge the ones he approved into reusable skills. Han runs skills inside controlled environments, grades the result, and preserves the complete trajectory so we can inspect what the agent actually did. Preserve the intent. Inspect the trajectory. Verify the result. Turn what works into skills. Full episode in the replies 👇

• 
  Joey Chang (@iwhaleocean) reported 23 minutes ago

  Another failure: the first plausible root cause won. For a GitHub OAuth 400, one path blamed the proxy; later evidence pointed to Supabase configuration. Now separate agents test code, config, upstream, and environment hypotheses before anyone proposes a fix.

• 
  Moysei (@0xMoysei) reported 24 minutes ago

  8 small gold boxes on a desk just ran a 440GB frontier model with 400,000 tokens of context. A datacenter doing the same job with GPUs would bill $150,000. This desk cost $36,000. The video shows the whole heist. 8 NVIDIA DGX Sparks wired into one brain: 1 terabyte of combined memory, 160 ARM cores, a 400G switch split into 100G lanes feeding every node. The model is GLM-5.2, full weight, no cut-down version. It loads sharded across all 8 boxes and answers at 24-26 tokens per second, with prefill near 880 tokens per second holding all the way to 300K context. He ran entire workdays on it before showing anyone. The comparison that breaks brains: matching this memory with RTX Pro 6000s takes 10 cards at $15,000 each. Same model, same quant, 4x the price, and the prefill chokes on CPU offload anyway. The desk wins on cost and holds its own on speed. Nothing worked out of the box. He patched vLLM by hand, forced static IPs across 8 nodes, tuned jumbo frames, and debugged alongside Claude Fable until 400K context ran stable. Every script is free on GitHub. Then the ending: the cluster ships to a datacenter next week. The desk was just its first job. $150,000 of capability, assembled where the coffee mug goes. The gap between "home setup" and "research lab" is now 8 boxes and a switch.