August 15, 2026: Cloudflare Daily: MCP visibility before Q2 answers

Welcome to Cloudflare Daily. It is August 15, 2026, and the most consequential new Cloudflare evidence is technical rather than financial. Cloudflare is pushing Zero Trust controls into Model Context Protocol traffic, the plumbing many AI agents use to call tools, APIs, and software services. On August 14, the company said Cloudflare One can identify inspected MCP traffic, show which users and servers are generating it, and control direct connections on managed network paths. The developer changelog adds the operational detail: Gateway now detects MCP traffic, exposes a beta HTTP policy selector called experimental dot is underscore mcp, and adds an AI security report showing MCP request volume, users, servers, and policies that target that traffic.

Why lead there? Because the enterprise risk around AI agents is shifting from what a chatbot says to what an agent can do. If an employee connects an agent to a database, code tool, ticketing system, or SaaS application, the agent may act with the employee's permissions. Cloudflare's blog argues that those actions can happen faster and less predictably than human clicks. That is a company framing, not independent proof of customer impact. But the product move is real enough to track because it sits at the intersection of Cloudflare One, Gateway, Access, SASE, and the Workers developer ecosystem.

A plain-language networking mechanic helps here. MCP traffic does not have to announce itself with a standard hostname or a path that says MCP. It can look like ordinary HTTPS API traffic. A gateway can only classify that traffic if it has signals to inspect, such as protocol headers or request patterns. If traffic is encrypted and the security gateway is not allowed to inspect it, visibility can be limited. That is also why Cloudflare's separate August 13 package-registry changelog matters. Gateway can now detect downloads from registries such as npm, PyPI, RubyGems, Cargo, Go, Maven, and NuGet using package selectors, but Cloudflare says TLS decryption is required for that registry security. In plain English, the guard post needs to see enough of the package request to enforce a package policy.

The important distinction is that availability is not adoption. The MCP selector is beta. The dashboard and controls can make AI-agent traffic more visible for administrators, but Cloudflare has not disclosed customer adoption, paid conversion, annual recurring revenue, or margin contribution from these features. The same caution applies to Workers AI. The August 14 changelog says DeepSeek V4 Pro and DeepSeek V4 Flash are available on Workers AI with a stated 1,048,576 token context window and require either a Workers Paid plan or prepaid AI Gateway credits. That is a pricing and availability signal, not evidence of usage volume.

There was also an adjacent Access update. Cloudflare says administrators can now protect a Worker across all associated domains and preview URLs, or make all new and existing Workers private by default with Cloudflare Access. That reduces a practical governance problem: developers can publish services in more than one place, and security policy can drift if each URL has to be protected separately. Again, no adoption number is attached, but the direction is consistent. Cloudflare is trying to make developer platform traffic and Zero Trust policy feel like one control plane.

The reliability record over the last two days is more mixed, though the disclosed incidents were all marked minor and resolved. On August 14, Cloudflare reported a Durable Objects and Workflows availability drop from Miami based colocations between 19:27 and 20:05 UTC. Earlier that day, it reported increased network congestion between eastern United States colocations from 15:34 to 16:28 UTC, with possible increased errors and latency reaching R2 and Durable Objects. Cloudflare also disclosed increased HTTP 5xx errors in Kuwait, Bangkok, Jakarta, and Dammam from 04:55 to 05:37 UTC. On August 13, it reported increased 503s when making configuration changes for Magic Transit, Cloudflare WAN, and Cloudflare One Appliance from 19:51 to 21:16 UTC, while stating that traffic was not impacted and only configuration changes were affected. These pages do not establish a global outage, customer count, revenue effect, or root cause.

For the operating scorecard, the cleanest confirmed baseline in this run remains Q1 2026. Cloudflare reported Q1 revenue of 639.8 million dollars, up 34 percent year over year. It reported a GAAP operating loss of 62.0 million dollars, non-GAAP operating income of 73.1 million dollars, free cash flow of 84.1 million dollars, and cash, cash equivalents, and available for sale securities of 4.1639 billion dollars at March 31. Do not mix those GAAP and non-GAAP numbers. They answer different accounting questions. Cloudflare also said in Q1 that it expected to reduce its workforce by about 1,100 people and incur 140 million to 150 million dollars of related charges, with most charges expected in Q2 and substantial completion expected by the end of Q3.

There is a housekeeping correction for today. Search results and social posts point to Q2 numbers, and Cloudflare had scheduled Q2 results for August 6. But in this run, I did not confirm a current primary Cloudflare or SEC Q2 financial package. Those unverified Q2 snippets are excluded. A failed SEC monitor capture is not evidence that no filing exists. It means the next episode needs the actual 8-K, 10-Q, earnings exhibit, supplemental, or transcript before updating the financial scorecard.

One finance mechanic to keep in mind is remaining performance obligations, or RPO. RPO is contracted value that the company expects to recognize as revenue later if it delivers the promised service. Current RPO is the portion expected within the next twelve months. It is useful backlog context, but it is not revenue today, not cash, not bookings, and not a guaranteed renewal. Cloudflare highlighted 34 percent current RPO growth in Q1. The follow-up question is whether AI, SASE, and large customer contract momentum turns into recognized revenue, margins, and free cash flow.

Commercially, Cloudflare's investor homepage says 45 percent of the Fortune 500 were paying customers as of June 30, 2026, and lists more than 4,600 large customers as of Q2 2026. Those are company claims and useful scale markers. They do not disclose contract value, product mix, retention, or whether the new MCP controls are driving wins. Customer stories on the website can identify use cases, but they are marketing and participant claims unless backed by independent evidence or filed data.

Finally, governance remains an open thread. The June 30 8-K says stockholders approved amendments to establish Class C common stock, implement the Class C split, increase authorized Class A and preferred shares, and require approval from a majority of independent directors for certain acquisitions using more than 100 million dollars of Class C stock consideration. That is a baseline for future acquisition financing analysis, not proof that a specific deal is coming.

Here is what to watch next. First, a primary Q2 financial package: the 8-K, 10-Q, exhibit, supplemental, or transcript. Second, any MCP security evidence beyond launch, such as named customers, pricing, dashboard adoption, or earnings-call commentary. Third, postmortems or repeat incidents involving Durable Objects, Workflows, R2, Magic Transit, Cloudflare WAN, or Cloudflare One Appliance. Fourth, actual Q2 restructuring charges, headcount disclosure, cash severance, stock compensation effects, and free cash flow. Fifth, any amended charter, equity-plan exhibit, or acquisition disclosure showing how the approved Class C structure is implemented. Cloudflare is scheduled for the Stifel Tech Executive Summit on August 24 and Goldman Sachs Communacopia plus Technology Conference on September 9, so those are falsifiable dates for new management commentary.