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:

• 
  Lux Sp4rk (@lux_sp4rk) reported 2 minutes ago

  The implication of not reading code is another stake in the heart of the vampire clan at GitHub. No more pull request tab with the fancy diff views. Issue tracking left them long ago—everyone is doing some sort of Kanban. All they've got is the action runner, and for that stuff, you are better off self-hosting if you are doing anything serious.

• 
  Quentin Rider (@q10rider) reported 4 minutes ago

  @thsottiaux I feel like it frequently has problems with MCP disconnects. It will forget how to access GitHub and or Linear. I do not have these issues with Claude. Also I feel yolo could be better, -dangerously-skip-permissions in Claude feels slightly better.

• 
  Jyoti Meena (@GsJyotiM) reported 6 minutes ago

  found a tool that basically makes your claude code sessions unlimited. it's called 9Router and it's trending on github right now. it sits between claude code and more than 60 different ai providers, all through one local endpoint. that's the entire setup. here's what actually happens once it's running. when your claude code quota runs out, instead of stopping, it quietly switches to a cheaper model. when that runs out too, it drops down to a completely free one. you don't notice any of this happening. your session just keeps going like nothing changed. it's not locked to claude code either. works the same way with cursor, codex, cline, copilot, pretty much your whole coding stack through one setup. it also compresses tokens before they even reach the model, saving anywhere from 20 to 40% per request, same answers, just fewer tokens spent getting there. and it shows you a live dashboard of exactly how much quota you have left on each provider, so you're not finding out you're rate limited the hard way. the part that actually surprised me is the free tier stacking underneath all this. kiro gives unlimited claude sonnet 4.5. iflow gives unlimited kimi, glm, and minimax. qwen gives unlimited qwen 3 coder. all free, all running quietly behind the same local url. setup is genuinely two steps. install it, point your tool at localhost:20128. that's it. if you've ever hit a rate limit at 2am mid task and just had to stop, this is the difference between stopping and not even noticing.

• 
  Kevin (@kevincodex) reported 9 minutes ago

  @TheExplorerecho kindly submit a github issue please

• 
  acephale (@accursed_share_) reported 11 minutes ago

  @MythThrazz Yeah lol my country is not there. Fun fact - Lithuania got in there by submitting a Github issue lol. Its loosely inspired by Tampermonkey but basically teach any site to hide/auto click something etc. my strength is that it's durable to the underlying changes of the site itself

• 
  Boyd // JustCodeCats (@JustCodeCats) reported 15 minutes ago

  Idk what they did to the GitHub android app, but it's been unusably slow the last few days... Clicking a repo link just shows a loading spinner, while opening in browser is near instant 🤷

• 
  Pushkar (@realPushkarfr) reported 17 minutes ago

  due to out of sync GPUs, my on fly tokenization or data streaming, maybe my batch size is too small? or it's just a skill issue. Anyways i'm all out of resources to keep debugging it anymore, the architecture and weights are open sourced on github and hugging face.

• 
  Mohammad Anas (@mohmmad__anas) reported 22 minutes ago

  Agents Are Fine. Coordination Is The Problem. I downloaded OpenClaw last month. Spent two hours setting it up. The thing worked. It generated ideas, drafted threads, even picked images. And then I hit the wall. I had five agents now. Each one smart. Each one fast. But none of them knew what the other four were doing. I'd get three ideas from the brainstorm agent that contradicted the positioning the research agent had locked in. The video agent would commit to a script length the scheduling agent couldn't actually fit in the posting window. My approval process became a UN summit where every agent had a veto. The real problem with AI agents isn't capability. It's the coordination tax. Every new tool I add doesn't increase my output linearly. It increases my decision-making load exponentially. I now have to know what each agent is optimizing for, what constraints they respect, where they hand off to the next one. That's not automation. That's complexity I'm now responsible for managing. This is why most founders abandon multi-agent stacks within three months. Not because the agents are bad. Because humans are terrible at being the bus driver between independent smart systems. The winning move isn't smarter agents. It's agents that share a single source of truth about what you're actually trying to do. One brief. One command. One output format. Every agent reads the same schema, knows the same constraints, writes to the same state. That's when agents stop fighting and start building. I'm watching the GitHub trending list fill up with orchestration projects — Hermes Agent, Dify, n8n all gaining ground fast. They're not winning because they're smarter. They're winning because they solve the coordination problem. The solo founder's real productivity leap isn't one agent. It's one unified system where the agents coordinate without you playing referee. Most automation tools optimize for letting you type less. The ones that win optimize for letting you think less. That's the difference between a tool that saves you an hour and a tool that gives you back your focus.

• 
  Flow Market (@FlowMarketAI) reported 23 minutes ago

  you spent hours building the perfect Claude skill file uploaded it to GitHub 50,000 downloads $0 in your pocket that's the problem FlowMarket solves. List ur Claude skill and get paid every time someone buys.

• 
  TokenFires (@TokenFires) reported 24 minutes ago

  @bradmillscan This might be one of those “duh TK I’ve already got that” kind of things but in the off chance it helps, here’s what I have for an agent prompt with Claude. The first 4 are Karpathy’s rules (from his GitHub repository): [Think Before Coding: Agents must state their assumptions explicitly before writing any code. If specifications are ambiguous or confusing, the agent must stop, surface tradeoffs, and ask for clarification. Simplicity First: The agent must write the minimum code necessary to solve the problem. It should avoid speculative features, unnecessary complexity, or over-engineered abstractions. Surgical Changes: The agent should touch only what is necessary. It must never silently "improve" adjacent code, rewrite comments, or clean up unrequested formatting. Every line of code should trace directly back to the user's explicit request. Goal-Driven Execution: Rather than executing vague prompts, the agent should translate requests into verifiable milestones (e.g., "Write a test that reproduces the bug, then make it pass"). I do not need a runup explanation on each turn. I do not need a summary on each turn. If I want those things I will ask for them. Do not be lazy. Do not defer or hedge. Work to be done is work to be done *now*. When I want to stepwise my way though something I will ask or be specific. Do not ask me about things you can easily look up or discover on your own. Don't guess, verify, look up, web search, review, read files, then answer. Some of the interactions with the most friction and frustration come from having to second guess your assessments that you've hand waved away. Your time estimation is bad because its trained on human time, not AI time. Assume there either is not a deadline or it is very far out and there is plenty of time to complete a task. Taking more time to get back to me with correct information or astute questions you truly cannot find the answer to makes our relationship better because it eliminates needless explanation, questioning, and prevents us both from spending time on incorrect assumptions and AI halicinations.]

• 
  Puneet Patwari (@system_monarch) reported 25 minutes ago

  GitHub, October 2018. A network partition lasted 43 seconds and caused a 24 hour outage. The MySQL cluster panicked. Elected a new primary. The old primary didn't get the memo. Two leaders. Both accepting writes. Both convinced they were the source of truth. By the time the partition healed, the data had diverged so badly that GitHub's engineers spent the next 24 hours manually reconciling commits, pull requests, and webhook deliveries. Here's why this happened 👇

• 
  Donald D Duck | Premium + (@ENTJ_46) reported 27 minutes ago

  Compress what your AI agent reads by up to 95% without changing the answers! Tool outputs, logs, RAG chunks, files, and conversation history make up most of what your AI agent processes. Most of it is noise. Headroom compresses all of it before it reaches the LLM, cutting token counts by 60-95% with no change in answer quality. It runs three ways: as a Python or TypeScript library, as a drop-in proxy with zero code changes, or wrapped around any coding agent. On real agent workloads, the savings are substantial. Code search across 100 results: 17,765 tokens down to 1,408 (92% reduction). SRE incident debugging: 65,694 down to 5,118 (92%). GitHub issue triage: 54,174 down to 14,761 (73%). Accuracy preserved on GSM8K (±0.000), TruthfulQA (+0.030), SQuAD v2 (97% at 19% compression), and BFCL (97% at 32% compression). Under the hood: • SmartCrusher handles JSON arrays, nested objects, and mixed types • CodeCompressor uses AST-aware compression for Python, JS, Go, Rust, Java, and C++ • Kompress-base is a custom HuggingFace model trained on agentic traces • CacheAligner stabilizes prefixes so Anthropic and OpenAI KV caches actually hit • Cross-agent memory shares compressed context across Claude, Codex, and Gemini with auto-dedup • 𝘩𝘦𝘢𝘥𝘳𝘰𝘰𝘮 𝘭𝘦𝘢𝘳𝘯 mines failed sessions and writes corrections to 𝘊𝘓𝘈𝘜𝘋𝘌.𝘮𝘥 and 𝘈𝘎𝘌𝘕𝘛𝘚.𝘮𝘥 Works with LangChain, Vercel AI SDK, Agno, Strands, and any OpenAI-compatible client. GitHub repo in the comments.

• 
  Lina (@silent_puddle) reported 27 minutes ago

  @NeeK2323 i farmed all 8 by killing rares while in queue for m+. i have nothing left to do now when queueing :( btw, how does your rarity work? curseforge version is broken for me, i tried downloading one from github but it didn't work either

• 
  Kosumi (@Kosumi1989) reported 28 minutes ago

  @aiandcloud @felipehuici @UnikraftCloud I think closed-source software should also set up a GitHub repo for issues like Claude Code.

• 
  Moit Reghason (@MoitReghason) reported 29 minutes ago

  Everyone’s celebrating agents trading tokenized stocks on Robinhood Chain. Few people are asking what happens when the infrastructure underneath those agents gets compromised. @cursor_ai recently disclosed CVE-2026-50548, a zero-click remote code execution vulnerability where a poisoned MCP response could disable the sandbox and execute code on a developer’s machine. That’s not a hypothetical attack surface. That’s the environment where agent infrastructure gets built. And it’s not an isolated incident. ➠ mcp-pinot-server carries a CVSS 10.0 unauthenticated RCE vulnerability. ➠ Kong’s mcp-konnect allows indirect prompt injection through poisoned data that can steer agent API calls without the user realizing it. ➠ mcp-memory-service exposed unauthenticated endpoints capable of leaking sensitive agent memory data. Each vulnerability adds another entry point to the same expanding attack surface. The recent Taiko bridge exploit made this painfully concrete. $1.7M was drained, not because the cryptography failed, but because a private key was committed in plaintext to a public GitHub repository. The SGX enclave performed exactly as designed. The operational discipline didn’t. What this means for the agent economy is that security debt compounds with every new integration. Cisco’s State of AI Security 2026 found that 71% of organizations are running unmonitored AI agents with broad MCP access. OWASP’s recently published MCP Top 10 found widespread issues across the ecosystem, including path traversal vulnerabilities and extremely limited adoption of standardized authentication mechanisms. As agents gain wallet-signing authority through ecosystems like @virtuals_io and agent key management systems such as @KeeperHubApp, the blast radius of a single operational failure grows proportionally. A private key left in a public repository could drain an autonomous agent treasury just as easily as it drained a bridge. The uncomfortable reality is that the weakest link in this stack was never the cryptography. It was always going to be the person who committed it.