An operating system for AI productivity — built by the team that built Workers, running inside Cloudflare every day, and open source so you can build your own.
SESSIONArchitecture · Security · Get Started
FORMAT90 min briefing + 30 min Q&A
AUDIENCESecurity · Cloud · Leadership
01
Cloudflare
The connectivity cloud — one network, every layer.
GLOBAL SCALE
One network, one control plane.
210+ cities for AI inference
80% of top 50 genAI companies
215B threats blocked / day
330+
Cities
125+
Countries
13,000+
Network interconnections
<50ms to 95%
of the world's Internet users
THE COMPLETE PORTFOLIO
Connect. Protect. Build.
Cloudflare's connectivity cloud — one control plane across security, performance, and developer services.
click the diagram to step through the four components →
PILLAR 04 · AI SECURITY / COST / PERFORMANCE
AI Gateway
One control plane between your apps and any LLM — for security, cost governance, and performance.
Guardrails & DLP — security
Caching & rate limiting — cost
Dynamic routing — performance
Full request logging — visibility
Solves: Token cost + Shadow AI visibility + model lock-in
PILLAR 04 · AI SECURITY / COST / PERFORMANCE
AI Gateway — the control plane for every LLM call.
Apps & Agents
Skills · harness · deployed agents
↓
Cloudflare AI Controls
Cloudflare's control plane for AI traffic
AI GatewayMCP Server PortalsZero Trust Network Access
Cloudflare Workers AI
External Inference
Cost-efficient AI usage for every employee:
Model optionality
Control model usage without lock-in risk.
Dynamic routing
Map employee roles and skill files to specific models.
Faster inference
Faster, more cost-efficient inference.
Caching
Save token spend on at-scale usage.
PILLAR 03 · PROTECT AI-POWERED APPS
AI Security for Apps
Formerly Firewall for AI — a WAF-integrated detection layer for your own AI-powered, customer-facing apps.
LLM endpoint discovery
Prompt injection score (1–99)
PII detection, 37 categories
Unsafe topic detection
Model-agnostic, runs inline at the edge — the bidirectional counterpart to AI Gateway's outbound DLP.
PILLAR 02 · PROTECT AGENTIC AI ACCESS
Remote MCP Servers
Host MCP remotely — instead of locally on developer devices.
No ambient credentials on laptops
Centrally managed & monitored
Smaller local attack surface
PILLAR 02 · PROTECT AGENTIC AI ACCESS
MCP Server Portals
A governed portal for internal services & SaaS apps — capability-based, not ambient.
Capability-based access
Human-in-the-loop approval
Full audit log
Gatekeepers approve asynchronously — agents keep working, humans review on their own time.
Solves: Dev-speed-vs-review + Shadow AI visibility
PILLAR 02 · PROTECT AGENTIC AI ACCESS
ZTNA (Access)
Enforce Zero Trust policies to authenticate and authorize every request.
The same Access product you already sell — now securing agent traffic too.
BACK TO WHERE WE STARTED
Every pain point, accounted for.
VISIBILITY
Shadow AI sprawl
↓
AI Gateway + MCP Portal
SECURITY
Speed outpaces review
↓
Gatekeepers
PRODUCTIVITY
Duplicate work
↓
Shared skill library
COST
Token spend
↓
AI Gateway
04
Developer Platform
What Cloudflare OS is actually built on.
FULL PORTFOLIO
Every primitive, one runtime.
Cloudflare OS uses a slice of this — the rest is available for whatever ST Engineering builds next.
COMPUTE
WorkersDynamic WorkersPagesContainersWorkers for Platforms
STATE & STORAGE
Durable ObjectsR2KVD1HyperdriveQueues
AI
AI GatewayWorkers AIVectorizeAI SearchAgents SDK
REALTIME & MEDIA
Browser RunRealtimeStreamImages
ORCHESTRATION
WorkflowsSandbox SDKSecrets StorePipelines
Verified against the cloudflare-os source repo — wrangler bindings, env.AI + AI Gateway billing code, and the README's own architecture notes.
Orange = confirmed used by Cloudflare OS today
FOUNDATION
Built on Workers, by the Workers team.
Cloudflare OS
Workers
Every Gadget runs in a Dynamic Worker
Durable Objects
Every workspace is its own DO
Dynamic Workers
Sandboxed, spun up on demand
Facets
Gatekeepers install into each workspace
THE HARDER QUESTION
Why here, not AWS / GCP / Azure?
Agents are a different workload than websites — they chain tool calls, wait on LLMs, and need to run close to whatever they're calling.
DIMENSION
HYPERSCALERS
CLOUDFLARE
Cold start
Lambda: 100–500ms · containers: seconds+
Workers: under 5ms
Billing model
Wall-clock time — billed while waiting
CPU time only — not billed while idle
Agent-service reach
Bedrock / Foundry: a handful of regions
330+ cities, edge-native
Security model
IAM + separate WAF/CDN vendor, bolted on
Zero Trust, Access, Gateway — same network
Honest caveat: hyperscalers still lead on deep managed ML training and GPU pipelines. Cloudflare's story is serving and securing agents at the edge — not training foundation models.
ARCHITECTURE, NOT ASSEMBLY
Purpose-built, not assembled from parts.
On traditional cloud, an agent means stitching together six services before writing any agent logic.
Compute service
Database
Message queue
Container runtime
Secrets manager
API gateway
On Cloudflare, it's one integrated framework:
Agents SDK
Durable Objects — state lives with the agent
Workflows — durable retries + human-in-the-loop
THE PRICING ARGUMENT
Pay for thinking, not for waiting.
Agents spend most of their time waiting — on the LLM, on a human approval, on a scheduled trigger.
Traditional VM / container alive 24 hours
24 hours
Cloudflare Durable Object alive 24 hours, hibernates when idle
10 seconds
An agent that runs 10 seconds of actual compute but stays "alive" for 24 hours costs 10 seconds — not a day of a server.
05
Let's bring it to market together.
Cloudflare brings the platform. You bring the customer.
WHO DOES WHAT
Cloudflare provides the toolkit; you bring expertise & white-glove services.
Cloudflare's own
Partner augmented
Comprehensive security & cost controls
AI Gateway
Zero Trust Network Access
MCP server hosting + portals
Full-stack developer platform
Storage
Compute
AI
Starter kit of skill & context files
Cloudflare's own
Partner augmented
Services to build central context layer
Work with customer to document skills & context
Implementation & support
Develop custom frontend
Integrate with internal systems
Roll out
FOUNDATIONAL WORK
Invest in a central context layer.
Email Alias
Magic inbox for work requests
Expert Tasks
Documented expert workflows
↓
Triage & Audit
AI agent operators curate the work
↓
Skill Repo
Curated skill files + context, centrally managed
↓
AI Agent Workspace
Surfaced and ready-to-use in Cloudflare OS
Defining the jobs-to-be-done as centrally managed skill files:
1
Document workflows
Tap experts to document workflows and create skill files.
2
Open a magic inbox
An email alias where everyone else can send work.
3
Staff it with agents
AI agent operators run the inbox as a service.
4
Curate into skills
Collect the use cases and convert them to skill files.
5
Surface in Cloudflare OS
Everything is ready-to-use for the whole team.
SKILLS VS CONTEXT
Two file types, one knowledge layer.
Skill File
Teaches an agent how to do one task — like a recipe.
When to use it — what triggers the task
Step-by-step, each step an action verb
Common mistakes to avoid + if/then logic
One page ideal · .md
Context File
Gives an agent background knowledge — like a glossary.
Overview — 2–3 sentences on the topic
Key definitions — terms, plain-language
Quick reference table — "if X, use Y"
One–two pages · .md
06
Get Started
Standing up your own instance on your NFR account.