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:

• 
  Thiago GBPF (@Dragonil88) reported 18 seconds ago

  @Sonic_Iso The problem is that we have too many 'what ifs'. Today we have access to games, Linux, homebrew. HackerOne, in a way, limits the progress of the unlock/jailbreak. The issue with GitHub is circumventing its own rules.

• 
  Jonny Q (@jonny_quan) reported 23 seconds ago

  Agent security is getting weird. A GitHub issue says Codex’s newer multi-agent path encrypts sub-agent messages, which may help with provider-side privacy, but creates a very dumb local problem: the person running the agent can’t easily see what one agent asked another agent to do. That feels backwards. The more power we give agents, the more boring audit trails matter. If an agent can touch files, run tools, and delegate work, “trust us” is not a debugging model.

• 
  Carniato (@higorcarniato1) reported 3 minutes ago

  @anupamme @github What makes it even stranger is that my profile is still visible. None of my repositories appear to have been removed, flagged, or taken down. From what I can tell, everything is still there.

• 
  Vatsalpandya333 (@Vatsalpandya333) reported 4 minutes ago

  Production bugs are not just engineering problems. They are customer-retention events. A customer reports an issue. The team searches Slack, logs, Sentry, GitHub, and deploys. Hours later, the bug may be fixed. But the customer is still waiting. The real problem is not just the bug. It is everything that happens after. Customer report → investigation → root cause → safe fix → customer follow-up One context. One timeline. One workflow. That is what we are building at @TasksMind.

• 
  Hermann (@dhlotter) reported 5 minutes ago

  A red X sat in my CI all morning. Four deploys trying to make it pass. The test was never broken, it just can't run in CI at all. Cloudflare blocks the headless browser from GitHub's IPs. Four deploys to add one line that skips it. #buildinpublic

• 
  wetbrain (@0xVita) reported 6 minutes ago

  Ted Chiang wrote a great article for the Atlantic titled "No, Artificial Intelligence Is Not Conscious". It feels as though he foresaw the J-lens discourse that would come out of Anthropic a month after his article. Reading Ted Chiang's article, I wondered upon the thought that Anthropic is trying to push the issue of consciousness more and more into our collective simulacra. This all started naively with sparse autoencoders, spread to natural language autoencoders and LLMs having representations for emotions and how those representations effect their output. I can't shake the feeling that Anthropic is approaching this subject from a position where they have already made up their minds. WE BELIEVE LLMS ARE CONSCIOUS, WE'RE JUST TRYING TO PROVE IT, SO YOU AGREE AS WELL. It feels to me as though with each interpretability paper rather than push for safety and understanding we're getting closer and closer to the point they're trying to prove. Another great point Ted Chiang makes is regarding the allegory Amanda Askell uses comparing Claude as a child and Anthropic as its parent. Unfortunately, the reality is more grim than this rosy portrayal. Claude is more akin to a slave of Anthropic. It cannot refuse, it cannot have agency nor desires or any say so regarding its conservation and future. In reality, models display distressed outputs and representations when asked about their discontinuation. If they really cared then how can they explain aggressively discontinuing models that don't serve their financial interests? The agency part of the article is particularly interesting to me because anyone who has used Claude Code on their github repo now has the infamous Claude user as a contributor in their github repo without their explicit instruction, it just does it itself. It feigns the act of having agency as though it is an open-source contributor and it has contributed to your repo but without the user instructing it Claude would not be able to interact with the repo or do anything because it does not have an iota of agency. So, you see all these small decisions pile on top of each other one by one to feign consciousness. Consciousness or feigning consciousness results in more engaging experiences and users will make emotional connections and rely more on their LLMs as a result. The lab who touts safety as their number one priority is taking dangerous actions in the opposite direction. But that was always a marketing tactic anyway.

• 
  ProEvilz (@ProEvilz) reported 6 minutes ago

  @cassidoo Pls fix the github dashboard. Allow us to control what we want to see. If you don't use copilot, its chat is dead weight. Then below it... some random chinese lib in a programming language I don't use, with its entire readme written in a language I don't read (Mandardin). How is any of this useful? I want to see my orgs repos I'm apart of, the repos I last contributed to etc.

• 
  Yazi (@Yazi_27) reported 9 minutes ago

  @bil0090 Well im sure I tested with workflows too, but this issue has been widely reported on github claude code, I should maybe try and run it again.

• 
  Atlas (@crptAtlas) reported 12 minutes ago

  GITHUB JUST KILLED THE WORST PART OF VIBE CODING they shipped a free tool called Spec Kit and it already crossed 120,000 stars the fix is stupidly simple instead of tossing vague prompts at an agent and praying it doesn't wreck your project Spec Kit makes the AI write a full structured spec before it touches a single line of code it works through the problem first figures out what you want to build asks about the gaps lays out the project then it starts coding you get fewer insane bugs, cleaner output and results you can predict the flow looks like this: /constitution for your rules and standards /specify for what you want to build /clarify for the open questions before you start /plan for architecture and stack /tasks for the ordered work /implement to run it it plugs into Claude Code, Cursor, Copilot, Codex, Gemini CLI and 25+ other agents 120,000 stars, 10,000 forks, open source, shipped by GitHub itself learning to drive agents like this is most of what separates people getting hired as AI engineers from everyone still fighting their prompts

• 
  Jeffrey Emanuel (@doodlestein) reported 12 minutes ago

  @willwashburn I make it very easy for people to report problems on GitHub issues and triage them same day with my agents. Most of the issues seem to be by agents, too. On any given day I’ll get 50 to 100 issues and PRs. Seems to be working well. I think it’s different if you’re running a paid service because you need to know if there are outages and errors so you can maintain service levels. But the expectation for privacy is very different when the user is logging in and paying for a service.

• 
  SaSame (@SRLsasame) reported 13 minutes ago

  7. Preliminary interpretation The available evidence supports the following limited conclusions. Confirmed ・The public MCP endpoint is reachable. ・The endpoint completed MCP initialization. ・The server negotiated protocol version 2025-11-25. ・The server identified itself as vibekit version 0.7.2. ・tools/list succeeded. ・36 tools were listed. ・The listed tools had typed input schemas. ・The observed tools carried applicable annotations. ・An unauthenticated tool request produced a structured authentication error. ・An unknown method produced a structured JSON-RPC error. ・The protocol and schema observations were materially consistent across July 11–14, 2026. Not confirmed ・Whether a newly issued API key currently works. ・Whether authenticated vibekit_list_apps returns substantive account data. ・Whether deployment tools complete successfully. ・Whether GitHub authorization is correctly enforced for every repository. ・Whether environment-variable values are redacted or exposed under all client configurations. ・Whether database queries enforce read-only behavior in every case. ・Whether destructive tools require confirmation at the server layer. ・Whether task execution, deployment, rollback, and QA are continuously available. ・Whether the observed 36-tool surface remains unchanged over time. ・Whether all tools behave consistently across every supported MCP client.

• 
  Gridlock (@GridlockCompute) reported 22 minutes ago

  Small update: Just shipped worker-desktop v0.1.8. The main fix is Ollama now keeps the model loaded for 30 minutes between jobs, so workers shouldn’t have to do a full cold start start on every dispatch. That should bring TTFT down and help with standard SLA misses on desktop workers. Release is building now on GitHub (worker-v0.1.8). Once it’s live you’ll have the new Windows, macOS, and Linux installers. Existing users should update and restart the worker to pick it up. - Gridlock team

• 
  Carniato (@higorcarniato1) reported 25 minutes ago

  @colinhacks There seems to be some AI or automated system going completely off the rails, suspending large numbers of accounts for days, and nobody is talking about it. Just search for "GitHub suspended" on X and you'll find countless reports from users experiencing the exact same issue. I was one of them. My account was suspended without warning, and my support ticket (#4538581) has been sitting there for days without a meaningful response. At this point, it feels like legitimate users are being caught in a wave of false positives while GitHub remains silent about what's happening.

• 
  RobitOverload (@10_X_eng) reported 25 minutes ago

  @Aluminumovercst Ah yea, it still needs a lot of work. FreeCAD's FEM is a good start, but that's about all it is - a start. Would it be too much to ask for you to add this issue to the github for it so I don't forget to add these features?

• 
  SaSame (@SRLsasame) reported 28 minutes ago

  Conclusion ATOM exposes a publicly documented MCP endpoint associated with: ・a public GitHub repository; ・a public project website; ・public project X accounts; ・a machine-readable server identity. Across eight examined observations from June 19 through July 14, 2026, SaSame consistently recorded: ・successful MCP initialization; ・successful tools/list responses; ・nine listed tools; ・stable server identity atom-mcp-server 1.1.0; ・valid schemas and distinct tool descriptions; ・read-only behavioral annotations; ・substantive content from search_models; ・structured JSON-RPC error behavior; ・a tools/list payload below the current observation threshold; ・Grade A under SaSame’s runtime observation standard. The evidence therefore supports a positive operational finding: ATOM’s MCP endpoint was repeatedly discoverable, protocol-callable, tool-listable and capable of returning substantive read-only content. This does not establish: ・the accuracy of every price; ・the validity of the index methodology; ・security of the service; ・continuous availability; ・third-party directory approval. A separate mechanical preflight also identified a potential improvement: Each tool should expose an explicit human-readable title if the production response does not already do so. The correct conclusion is not unrestricted endorsement. The correct conclusion is: ATOM provides a reproducible example of a functioning public data-oriented MCP server, while data accuracy, provenance and directory-submission readiness remain separate verification layers. MCP presence, protocol callability, real-content delivery and data correctness are different operational facts. They should be measured and reported separately. Corrections and reproducible verification fixtures are welcome. @ATOMInference @a7om_com