Cloudflare open-sourced Cloudflare OS on August 4, 2026 — an Apache 2.0 agentic workspace the company has run internally since May, when CEO Matthew Prince rolled it out to every employee. The trigger was a sales rep requesting production API keys to roughly 12 systems of record to ship a personal "SuperApp." That request forced a choice: hand out credentials or build governance infrastructure first. Cloudflare chose infrastructure.

The platform has three components. First, a browser-based agent workspace — no terminal required — that loads each session with company-curated context and skills, so employees stop re-explaining procedures and terminology from scratch. Second, Gatekeepers: individual Cloudflare Workers sitting between Cloudflare OS and each external system. A Gatekeeper understands the target API, enforces field-level masking, applies rate limits, and holds side-effecting operations (a PR merge, a billing change) until a human approves. Credentials never touch the agent or its generated code. Server-side logic runs in a Dynamic Worker with global outbound networking disabled; client code runs in a sandboxed browser frame. Third, an apps layer where any agent output — a doc, a query result, a dashboard — can be promoted to a live Cloudflare Worker backed by Durable Object state.

The architecture separates deterministic steps from model-driven steps by design. Workflows use code for the predictable sequence and invoke a model only where judgment is needed — on demand, on a schedule, or triggered by an event from a connected system. That keeps token spend predictable and audit trails clean.

Cloudflare OS separates deterministic code paths from model-invoked judgment, reducing AI hallucination and speeding review cycles.
FIG. 02 Cloudflare OS separates deterministic code paths from model-invoked judgment, reducing AI hallucination and speeding review cycles.

Before Cloudflare OS, the internal AI posture was cautious: chat interfaces and boilerplate-code assistance, nothing touching production. The engineering problem once agents arrived wasn't quality — AI made bad code faster than review could catch it. The fix was a shared context layer encoding Cloudflare's coding standards, review checklists, and known patterns, so agent output arrives pre-idiomatic.

Non-engineers got the workflow builder. Describe a job in natural language; the agent writes the code; run it on schedule or event trigger. The IT help desk now runs a morning agent that surfaces open ticket queues and metrics. A procurement bottleneck that would have taken "days of manual spreadsheet crawling" was diagnosed by a single Cloudflare OS report. A laptop-replacement tracking workflow built by one IT staffer was shared directly with a finance colleague in the Lisbon office — no ticket, no sprint cycle required.

The hardest architectural problem emerged from collaboration. MCP server access tells you which tools an agent can call; it says nothing about which underlying resources the agent observed during a session. When employees started sharing workspaces and apps, Cloudflare had to rebuild the access model: any shared artifact must reflect the viewer's own permissions, not the creator's. That gap drove a full v2 rewrite. Cloudflare OS is Apache 2.0, but the runtime lives on Cloudflare's edge. The open-source license gives you the code; the execution environment is still one vendor.

The CFO updated long-term operating margin targets from 20% to 30% and free cash flow targets from 25% to 30–35%+ citing Cloudflare OS productivity gains — a concrete benchmark for teams building the business case for their own internal agent rollouts.

Cloudflare's CFO raised operating margin and free cash flow targets following Cloudflare OS productivity gains.
FIG. 03 Cloudflare's CFO raised operating margin and free cash flow targets following Cloudflare OS productivity gains. — Cloudflare Q1 2026 earnings