Cloudflare's internal AI engineering stack is a company-wide control plane. It gives employees one approved way to reach models, company knowledge, tools, and review agents while keeping identity, cost, permissions, and audit data in a central system.
What is Cloudflare's internal AI engineering stack?
Cloudflare's operating account describes a standard setup built around OpenCode, a proxy Worker, Cloudflare Access, AI Gateway, Backstage, repository guidance, Model Context Protocol tools, and AI code review.
An employee runs one bootstrap command. Requests pass through company identity and the central gateway. The gateway selects approved models, applies permissions, records usage, and avoids placing long-lived provider API keys on laptops. Teams can still use several model providers behind the same boundary.
This is a control plane: a shared place to manage who can use a capability, what it can reach, how it is observed, and what it costs.
How does the agent learn about Cloudflare?
Cloudflare uses Backstage as a knowledge graph for services, repositories, owners, and dependencies. Its April account reported 2,055 services and related catalog data. The company also generates repository instruction files called AGENTS.md across roughly 3,900 repositories.
The two context layers answer different questions. Backstage explains how systems connect across the company. Repository guidance explains how work should happen in a particular codebase. An agent can receive relevant structure without loading the whole organization into one prompt.
Cloudflare reported 27 internal MCP servers by August 2026, up from 13 in an earlier version of its portal. MCP, or Model Context Protocol, is a common way for an agent to call approved tools. The initial catalog was read-only. WriteGuard added a central, audited path for actions that change outside systems and was still in private beta when checked.
How does Cloudflare review agent-assisted code?
Cloudflare's AI code-review system can dispatch as many as seven specialist reviewers based on a change's risk. Review areas include security, performance, code quality, documentation, release readiness, and compliance. The system returns structured findings and can block serious issues.
Cloudflare reported use across 5,169 repositories during one month in 2026. That is deployment scale, not proof that every generated finding is correct. Public material does not provide a complete precision, recall, false-positive, defect, or reviewer-time series.
What adoption has Cloudflare reported?
Cloudflare's April article reported 3,683 active users of internal AI coding tools, about 60 percent of the company and 93 percent of research and development. It also reported 47.95 million AI requests across 295 teams. The article calls these figures “last 30 days” in its introduction, while the pictured dashboard is labeled February 5 through April 15. SWFT therefore treats the scale as company-reported and the exact measurement window as unclear.
The same article says the four-week average for merge requests rose from roughly 5,600 per week to more than 8,700. Cloudflare presents these figures together as operating evidence. The public article does not establish that AI caused the increase, or whether the additional changes improved customer outcomes.
Activity, adoption, accepted output, quality, and product value are separate measures. A control plane can make the first three visible. The factory still needs defect, rework, incident, review-attention, and customer-outcome measures.
Why are central controls useful?
Without a shared boundary, each team may distribute provider keys, invent its own model routes, connect tools with different permissions, and collect incompatible logs. Centralizing those responsibilities makes the system easier to audit and change.
The tradeoff is platform responsibility. A central gateway or catalog can become a bottleneck or a large failure boundary. It needs clear ownership, reliable service, privacy rules, safe defaults, and a path for teams to add capabilities without bypassing the control plane.
What can a product team copy?
Begin with one approved front door. Bind model access to employee identity. Keep provider credentials off laptops. Record cost and activity at a useful team level. Connect a service catalog and repository-local instructions. Start outside-system tools as read-only, then add reviewed write capabilities with narrow authorization and audit. Apply specialist review according to risk.
Cloudflare's case is unusually detailed, but all adoption and output figures are company self-reports. They do not replace an independent productivity or quality study.