Information technology / Cloud infrastructure
Cloudflare: Running AI coding tools across the internal engineering organisation
- Company
- Cloudflare
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
Cloudflare is an internet infrastructure company with roughly 6,100 employees. The company said that handing out AI coding tools is not enough to connect them properly to engineering work. The issue was that AI agents do not become useful without structured data about internal services, repositories and release history. The cost of model usage and the question of internal material being retained by outside model providers had to be solved at the same time. The source described the work as an eleven month effort to rethink not just how code gets written, but how it gets reviewed, how standards are enforced, and how changes ship safely.
Technology and data
The company pulled people from across the business into a tiger team called iMARS and built an internal AI engineering environment on its own platform. A single internal portal aggregates 13 production MCP servers exposing more than 182 tools across Backstage, GitLab, Jira, Sentry, Elasticsearch, Prometheus, Google Workspace and an internal Release Manager. The 34 tools on the GitLab MCP server consumed roughly 15,000 tokens of context per request, and the source said Code Mode reduced them to two portal tools. Cloudflare Access handles sign in and requests pass through a single proxy Worker to AI Gateway. Employee email addresses are mapped to a UUID stored in D1 with a KV read cache, so AI Gateway only ever sees the UUID. An hourly cron refreshes the model catalogue in Workers KV and injects a store value of false so that data is not retained. An AGENTS.md file is generated in each repository, and the service catalogue is filled by a self hosted Backstage.
Results
The company said that in the last 30 days 93% of its research and development organisation used AI coding tools built on its own platform. It said actual users numbered 3,683, which is 60% of approximately 6,100 total employees. It said the four week rolling average of merge requests climbed from about 5,600 per week to more than 8,700, and the week of 23 March reached 10,952, nearly double the fourth quarter baseline. It said the OpenCode AI Gateway view routes 688,460 requests and 10.57 billion tokens per day to four providers through one endpoint. On Kimi K2.5, a 256k context model launched on Workers AI in March 2026, one security agent processes more than 7 billion tokens per day, which the company estimated would cost 2.4 million dollars per year on a mid tier proprietary model and is 77% cheaper on Workers AI.
Limits and open questions
The source gives the company wide user share and the growth in merge requests together but does not separate how much of that growth came from the AI tools. It also gives no figures for changes in code quality or defect rates.
Sources
- The AI engineering stack we built internally — on the platform we shipblog.cloudflare.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.