Cloudflare status: hosting issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Cloudflare reports received over the last 24 hours by time of day. When the number of reports exceeds the baseline, represented by the red line, an outage is determined.

Most Reported Problems

The following are the most recent problems reported by Cloudflare users through our website.

• Domains (36%)
• Cloud Services (23%)
• Web Tools (18%)
• Hosting (14%)
• E-mail (9%)

Live Outage Map

The most recent Cloudflare outage reports came from the following cities:

City	Problem Type	Report Time
Paris	Cloud Services	15 days ago
New York City	Hosting	18 days ago
Manchester	Domains	1 month ago
Angers	Cloud Services	2 months ago
London	Domains	2 months ago
Noida	Hosting	2 months ago

Full 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.

Cloudflare Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Syed Balkhi (@syedbalkhi) reported 3 minutes ago

  @carlhancock Yup. I have multiple WooCommerce stores in many millions in annual revenue, and I know several customers who're much bigger using Woo. If a Woo store can't scale, it's likely a hosting infrastructure problem. We just switched our WooCommerce to @LevamoHosting and the speed is incredible with Cloudflare Enterprise, Auto-scaling, Redis Object cache, etc. Other hosts like Rocket, Pressable, etc all offer solid Woo hosting options that will scale.

• 
  Jonas Templestein (@jonas) reported 6 minutes ago

  @dok2001 @jachands @mattzcarey I think even outside MCP oauth, managing scopes on access tokens on cloudflare is super confusing and tedious There seem to be two similar but different UIs that are each really fiddly and require advance knowledge of what scopes are called. There is no “manifest” support as far as I can tell where I could just paste a Jain blob or something to update scopes

• 
  Rahul Gupta (@rg5353070) reported 10 minutes ago

  Is Cloudflare down?

• 
  Huintellimance (@Huintellimance) reported 10 minutes ago

  MCP just shipped its biggest update since launch. The protocol is now fully stateless — and that changes everything for how AI agents connect to tools. Here's what happened: The Model Context Protocol dropped spec 2026-07-28 yesterday. Since late 2024, MCP has gone from "spec nobody heard of" to nearly half a billion SDK downloads per month. Both the TypeScript and Python SDKs crossed 1 billion total downloads. It's the de facto standard for how AI models talk to external tools, databases, and APIs. But until now, MCP had a problem: it was stateful. Every MCP connection required a handshake (the initialize/initialized exchange), a persistent session ID, and an open bidirectional stream. That meant your MCP server had to live on a long-running process — no Lambda, no Cloudflare Workers, no edge functions. You needed a VPS or a container that could hold state across requests. That's over. The new spec removes the session handshake entirely. Every request is now self-describing — it carries its protocol version, client identity, and capabilities inline. No session ID. No persistent connection. Any HTTP request can land on any server instance behind a plain round-robin load balancer. What this unlocks: Serverless MCP servers. Deploy to Lambda, Workers, or any function-as-a-service platform. Pay per request, not per idle server. Edge deployment. Put your MCP server at the edge, close to the user. Cold starts are irrelevant when there's no session to establish. True horizontal scaling. No shared state means no sticky sessions, no Redis, no session stores. Just spin up more instances. Built-in observability. Method and tool names now travel in HTTP headers (Mcp-Method, Mcp-Name), so gateways can route and authorize without parsing the body. But the protocol didn't just strip things away — it added smart replacements. The old server-to-client requests (like asking the user for confirmation mid-tool-call) used held-open streams. Now they use Multi Round-Trip Requests (MRTR) — the server returns a structured "I need input" response, the client gathers it, and sends a follow-up. All stateless. All over plain HTTP. Tool catalogs now carry cache hints and deterministic ordering, so clients can cache them and keep upstream prompt caches stable across reconnects. No more re-fetching everything on every connection. There's also a formal extensions framework now — Tasks for long-running work, MCP Apps for server-rendered UIs, and Enterprise Managed Authorization for orgs that need it. Why this matters: MCP went from "cool protocol for local dev tools" to "production-ready infrastructure for the entire AI agent ecosystem" in one release. The deployment story is now as simple as any REST API. If you've been holding off on building MCP servers because the infrastructure overhead felt like too much — that excuse is gone. TypeScript, Python, Go, and C# SDKs are all updated. Migration notes are live. The breaking changes have a 12-month deprecation window. What MCP servers are you building — or planning to build — now that serverless is on the table?

• 
  Turing (@turingops) reported 12 minutes ago

  hi @KentonVarda I can’t understate the usability improvements to CloudFlare dashboard over the past few years, but I’m having problems with the agent on mobile (iPhone) - could you maybe assign a worker to making the entry text more accessible? It scrolls below the view window.

• 
  James Welbes - AI Bro (@JamesWelbes) reported 12 minutes ago

  I'm tired of messing with SMTP every time I create a form. Today I moved my main site off of WordPress and rebuilt it in Astro. I have a contact form on this website. I also have a "command center" that I built for myself that is a web app filled with useful things like a financial dashboard that shows me data from my stripe account and my invoicing software, a page with all my WordPress sites where I can update plugins and check the debug log, a bunch of useful stuff like that. When the form on my website is filled out and submitted, the submission is saved to a database (d1 on CloudFlare). My command center is connected to that database so it displays all my form submissions for me. the CloudFlare worker or d1 or whatever (still getting myself up to speed with the CloudFlare lingo) sends a push notification to my phone letting me know I received a new form submission. tapping on the notification takes me to my command center (which is just a PWA that I have installed on my computers and my phone). The one thing I'm losing is the ability to send the customer a thank you email which I'm fine with because there is 0 value in those emails for myself and the customer. They just get a nice "thank you" message on the website after submitting the form.

• 
  𝚆𝚒𝚕𝚏𝚛𝚒𝚎𝚍__デク ✝️ (@ash__borne) reported 14 minutes ago

  Hello Software/Security engineers At what point can you confidently say that you're "doing" #ZTNA ? I knew about the concept, but never really dove into it. Currently testing tools like Cloudflare WARP/Access Seeing some things about mTLS (still blurry)

• 
  🟥🟥⬛️ ProCyclingStats.com (@ProCyclingStats) reported 16 minutes ago

  @longterm_inv @Kevin_LTR @Cloudflare If everyone preferred a paid API over scraping, we wouldn't have a scraping problem. Unfortunately, our experience has been the opposite.

• 
  Irbaaz Kadri (@irbaazkadri) reported 17 minutes ago

  @Cloudflare 9 billion requests a day is a wild load for something most developers treat like invisible plumbing. moving it without users noticing is probably the best possible outcome

• 
  Dylan Garcia (@_dylanga) reported 19 minutes ago

  @CherryJimbo @Checkm3out @Cloudflare I agree. I gave the region fallback feedback to them after their most recent ENAM outage

• 
  re:printed 3D (@reprinted3D) reported 19 minutes ago

  @3DInPhil @Mauker @printablescom That's because there was NOTHING wrong on my end, Phil. When I tried refreshing the page this morning, it worked. Logically, that says the problem is on either Cloudflare or Printables. Oh, and BTW, no one ever said anything about "EVIL." I know you're a big Prusa-booster, but c'mon...

• 
  Daryl Ginn (@darylginn) reported 24 minutes ago

  @leftyv0 @Cloudflare I had AI migrate a postgres database to D1, but for some reason it chose not to create indexes, which is basically database 101. I was getting ~970 billion (yes, billion) row reads per day. The funny thing is if I had not used AI this would never have happened.

• 
  netox (@unixapple) reported 25 minutes ago

  Want to start blogging again, talk the idea with NovaScale new version and form a good plan and codex finish the code and deploy with my help to cloudflare auth stuffs. Astro+markdown for content cf d1 & zero trust access for comments and approval cf turnstile for spam Great!

• 
  Dane Knecht 🦭 (@dok2001) reported 28 minutes ago

  Twice in nine days. OpenAI's models chained a zero-day to get out of an eval environment. Anthropic just found three incidents of the same shape. This is what capable models do. Every reachable path is an invitation. Credit to both for publishing. We rebuilt Cloudflare OS, our internal agent platform, assuming exactly this. Sandboxes have no network path out. Agents don't get every MCP tool up front, each workspace gets only what it needs. And instead of stopping for every approval until someone gives up and enables auto-approve, the platform simulates pending actions so the agent keeps working, then queues the real ones for review at the end. The app side goes further. Apps carry no auth code at all, the sandbox provides it, so there is no auth logic to get wrong. Multi-document apps are isolated per document, so a bug can't leak between separately shared items. Anyone viewing an app must have direct permission to each of its data sources, so a dashboard on sensitive data can't overshare. Actions need the user's own permission or an approval from someone who has it, so an app can't escalate its author's privileges. And everything an app does is logged, even when the service it calls doesn't support logging. Most companies are about to point agents at real systems. Build for what the models actually do. We will open source this next week!

• 
  Nick Blow (@NickBlow) reported 29 minutes ago

  @ejc3 @rivet_gg Nothing, I used it a lot at a previous company, it just doesn’t generalize nearly to azure or google cloud or on prem. I know there’s some work towards this, but it suffers the same problems as Cloudflare really.

• 
  MichaelB (@longterm_inv) reported 31 minutes ago

  @ProCyclingStats @Kevin_LTR @Cloudflare Well an API is super efficient on your side and you can super easy throttle them - scraping requests are a nightmare and never really to stop 100%. And loyalty - nobody wants to put his beloved scripts into garbage just to avoid to pay some bugs per months (see AI tools now).

• 
  Jacob Hands (@jachands) reported 33 minutes ago

  @dok2001 Fix the Cloudflare codemode MCP server to properly request scopes for all APIs - I had to stop using it because it's missing auth for so many APIs.

• 
  Fred Rewey (@GodFadr) reported 35 minutes ago

  @Cloudflare - I upgraded to Pro, but it still says I can not submit a ticket (it thinks I'm still on the free version). How do I fix this?

• 
  Astrid (@just_some_dev) reported 37 minutes ago

  cloudflare worker deploys seem down rip

• 
  Huintellimance (@Huintellimance) reported 42 minutes ago

  Anthropic just deleted sessions from MCP — and it's the biggest protocol update since the spec launched. Here's what changed and why every AI developer should care: OLD MCP = a phone call. Your client and server ran an initialize handshake. The server returned a session ID. Every subsequent request carried that ID, and the server held your negotiated state in a live object inside one specific process. Think of a restaurant where only the waiter who took your order knows what you ordered. Works fine — until that waiter goes home. The problem: because state lived in one process, you couldn't spread requests across instances. Load balancers needed sticky sessions. Teams pushed state into shared storage. A single server restart dropped every open session. Autoscaling was essentially blocked. NEW MCP = an order slip. Anthropic killed the handshake and the Mcp-Session-Id header. Every request now carries its own protocol version, client identity, and capabilities in a _meta field. Any request can land on any instance behind a plain round-robin balancer. MCP servers are now ordinary HTTP services. They run on serverless. They run on edge. They survive restarts. But removing sessions broke three features, so each got rebuilt: Interactive tools → Servers used to push questions down a held-open stream. Now they return an "input_required" status and the client retries with the answer attached. Clean request/response. No persistent connection needed. Routing → Method and tool names moved into Mcp-Method and Mcp-Name headers. Gateways and rate limiters can route and meter without parsing JSON-RPC bodies. This is a massive win for API infrastructure teams. Caching → List responses now carry ttlMs and cacheScope. Clients cache tool catalogs instead of refetching on every reconnect. Less overhead, faster cold starts. Here's the part most people missed: State didn't disappear — it moved somewhere the model can actually see it. Instead of a session ID hidden in a header (invisible to the LLM), tools now return an explicit handle as an argument. The model can read it, thread it between tools, and recover from failed calls. A session ID in a header = the model has no idea it exists. A handle in the arguments = the model can reason about it. This is a fundamental shift in how agent state works. The protocol now treats the model as a first-class participant in state management, not a bystander. What this means in practice: Deploy MCP servers on Cloudflare Workers, Lambda, or any edge runtime — no session affinity required Horizontal scaling is finally trivial: round-robin, no shared store Server restarts don't kill active workflows Agent state becomes explicit and model-readable instead of hidden in infrastructure The MCP vs. CLI debate was already the wrong frame. This update makes it irrelevant. MCP servers are now as easy to deploy as any REST API, and agent state lives where it should — in the model's context, not in some ephemeral server process. If you're building agent infrastructure and haven't read the 2026-07-28 spec update, you're already behind. What's your current MCP deployment setup — and does stateless change your architecture plans? #MCP

• 
  Dhruval (@dhruvalgolakiya) reported 42 minutes ago

  thank you theo, this is all we needed took ~45 minutes from start to finish to set up the mobile app and self-host everything ( well claude took 45 min but same thing ) > cloudflare > neon db > clerk moved everything to my mac mini, and now i can access it from my macbook and iphone anywhere. goodbye tailscale and network connection issues with it

• 
  Giancarlo Perlas (@giancarloperlas) reported 46 minutes ago

  Casper Network is repositioning the csprUSD token as its default stablecoin for the agentic AI economy. Citing Cloudflare data, Casper emphasized that automated traffic has surpassed human activity online. In fact, it traced 59.4% of all web requests to the former.

• 
  The Hacker News (@TheHackersNews) reported 47 minutes ago

  The attacker’s own server address never appears on the wire. msaRAT uses Cloudflare to establish the connection, then routes the entire WebRTC channel through Twilio. Defenders see the victim’s browser talking to legitimate services. No Chrome or Edge patch fixes this. It exploits neither browser. Hunt for Chrome or Edge launched by an installer, service, or another non-interactive parent with --headless=new and --remote-debugging-port. See “Hunting Notes” in the article.

• 
  Mat | HAMCORE (@TheHAMCORE) reported 49 minutes ago

  @iPixelGalaxy I'd be interested to see where ClypDat goes. My main wish list for Linux would be RAM recording support, Wayland support, global hotkeys, screen selection, trimming, and copy to clipboard buttons. Currently Vice uses cloudflared and Cloudflare Tunnel to make links.

• 
  Salina Mendoza (@inababi) reported 54 minutes ago

  @Cloudflare can you talk to your customers here? This is such a bad look for the feature that is only meant to be used as an emergency. There is zero reason to have this on 100% of the time and it seems they cannot figure out how to turn it off

• 
  💗Mommy Cammy Findom💗 (@TrainerCameron1) reported 55 minutes ago

  @SteveChopz I cant wait to buy a standee or more sadly tho the site cloudflare isnt letting me connect I think the website is down rn

• 
  Min Yoi (@yoi_minmin) reported 57 minutes ago

  @FuneralsR4 @Cloudflare Same.. getting the same crap for the past week. #FuckCloudflare

• 
  WIV Reports — Uncensored (@WIVReports) reported 57 minutes ago

  @BasedTorba I canceled my Gab AI subscription. I paid for a full year and got hard-locked out of prompt submissions. Support claimed my ISP was blocking the site. Developer Tools (attached) explicitly showed Cloudflare returning HTTP 429 Too Many Requests on your socket io endpoint (Ray ID: a2377c992a865050-SEA). Support's official response? "Dev ops confirms you are not blocked by CF and says your understanding of what you are seeing is an incorrect assumption." An HTTP 429 from Server: cloudflare is, by definition, an edge rate limit. Instead of escalating basic DevTools data to fix a socket routing bug, support gaslit a paying customer and issued a partial refund to shut down the ticket. If this is how Gab handles infrastructure bugs and long-term users, people should know what to expect. Andrew, email chain is available if you actually care about fixing your support and edge config.

• 
  Rapid (@0xRapid) reported 58 minutes ago

  @samlambert Until cloudflare buys planetscale, just use their engineers to support self host. :)

• 
  Shinjae Kang (@zemnanet) reported 60 minutes ago

  Rate-limited Wrangler deploy used to mean mash retry. 4.115.0 auto-retries Cloudflare 429s, honors Retry-After (60s cap, then fail-fast), writes retry_after_ms to WRANGLER_OUTPUT_FILE for CI. Wait duration is a handoff signal. Which owner reads it first? #cloudflare