Cloudflare cf CLI Hands AI Agents the Keys to 3,000+ API Calls

What Cloudflare shipped

Cloudflare has introduced cf, a new command-line tool designed from the outset to be operated by AI agents as much as by people. The company positions it as a single interface to its entire API, so that an agent can set up a Worker, deploy and monitor it, place it behind Cloudflare Access, buy a domain and put the Cloudflare WAF in front of it without leaving one tool 3. CX Today, covering the release as part of Cloudflare's Birthday Week 2026 announcements, puts the scope at more than 3,000 API operations 1.

The launch has four main pieces. There is a purpose-built command search and steering system that helps agents find the right command. JSON is the default output, pretty-printed for humans and condensed for agents to save context. A TypeScript configuration file, cloudflare.config.ts, starts with Workers. Vite becomes the default local development server 3. A DEV Community breakdown describes these together as a workflow that runs "from intent to API call to reviewed configuration" 2.

Why Cloudflare built it for agents first

Cloudflare's own usage data explains the design choices. According to CX Today, agents accounted for 25% of usage of Wrangler, Cloudflare's existing Workers CLI, in March, and that share had climbed to 48% shortly before cf launched 1. The same report notes Cloudflare's claim that automated traffic overtook human traffic in May 2026 and that daily AI-agent requests grew more than 1,700% year over year 1.

Cloudflare frames cf as a response to a broader trajectory in which agentic development is reshaping how software is built and deployed. It also says its command-discovery tooling is something it expects other CLIs to adopt 3. CX Today places cf within a wider Birthday Week push that included Forge, BEACON, end-to-end Traces and a feature called Issues, which routes Worker error context directly into coding agents or on-call workflows 1. In that framing, machine traffic is no longer a curiosity. It is operational traffic that needs dedicated controls 1.

Configuration as typed code

The most consequential change for day-to-day work may be cloudflare.config.ts. As the DEV Community analysis explains, a TypeScript file can expose a schema to an editor's language server. An agent editing a binding therefore gets completion and type feedback as it works, which is harder to do with Wrangler's TOML and JSONC formats even though those could be validated 2. The config uses a defineConfig function, and values can be computed from Vite's mode. That lets development and production share a common base instead of duplicating environment blocks 2. Cloudflare reportedly saw some internal configuration files of more than 5,000 lines shrink by about 40% using this pattern 2. Cloudflare says the format will eventually cover the rest of its platform beyond Workers 23.

The Vite default reflects history. Wrangler built its own bundling and dev experience before Vite existed, and cf now relies on Vite and its plugin ecosystem 23.

The security question

All three accounts focus on capability and convenience. None details the guardrails around that reach. That gap matters. A tool that lets an agent purchase domains, reconfigure firewalls and change access policies concentrates a great deal of authority in whatever credentials the agent holds. Typed configuration reduces one class of error, malformed or mistyped settings. It does not by itself address scoping, approval steps or recovery.

For security teams evaluating cf, the practical questions are operational ones:

• How are secrets handled when an agent is the operator?
• Can live logs be streamed for real-time debugging?
• How quickly can a bad deployment be reversed?

The answers will determine whether cf can replace Wrangler in production pipelines or stays an experimentation layer for now. Organizations should verify current feature coverage against their incident-response needs before migrating.

Our reading

cf is less a new CLI than a statement about who Cloudflare thinks its primary user is becoming. The jump from 25% to 48% agent usage of Wrangler in a matter of months 1 suggests the company is following demand rather than inventing it. The design choices are sensible for that audience: discoverable commands, compact JSON and schema-aware configuration. Pairing cf with Issues, which feeds production errors back to agents 1, points toward a closed loop in which agents deploy, observe and fix code.

That loop is only as safe as its weakest control. Broad API coverage is the headline feature. The test of cf's maturity will be whether permissions, credential handling and recovery tooling keep pace with what it lets agents do.