All Research
Cloudflare Strategy & Architecture
Sterling Synthesis · All sources merged

Sterling Synthesis · July 2026

The definitive Cloudflare strategy for Outsource Access

Everything from three separate AI research sessions, merged into one strategic document. 128+ capabilities, pricing, architecture, and a concrete roadmap for OA.

4 research sources merged 128+ platform capabilities $0 egress fees on R2 335+ edge cities

Unified Platform Overview

Every capability across all three research sessions. Source badges show which AI found each item. OA relevance rated High / Medium / Low.

How to read source badges

Sterling = Claude Playbook. Gemini = Google Gemini Deep Dive. ChatGPT = ChatGPT 5.6 Expert Guide. All = covered by multiple sources.

Build & Host Applications

The products you write apps on. This is OA's primary layer.

Workers — Serverless compute All

Upload code and it runs across 335+ cities. No servers to manage. Sub-millisecond cold starts thanks to V8 Isolates (thousands of apps share one OS process, each securely sandboxed). Over 3 million developers. 40%+ of Y Combinator's Winter 2025 cohort built on it.

Pricing: Free tier: 100K requests/day. Paid: $5/mo base includes 10M requests + 30M CPU-ms. Then $0.30/M requests, $0.02/M CPU-ms. No bandwidth charge.

Key insight (Gemini): Billed on CPU time, not wall time. A Worker that waits 10 seconds for an API but uses 10ms of actual CPU is billed for 10ms. Massive savings for AI agents that spend most time waiting.

OA relevance: High — the engine behind every app we build.

Pages — Website and web app hosting All

Deploy frontend sites with automatic CI/CD from Git. Supports server-side functions alongside static assets. Where the UI lives for all our dashboards.

OA relevance: High — every client dashboard, research page, and internal tool is on Pages.

D1 — Serverless SQL database All

SQLite-based. Store structured data: users, tasks, orders, permissions. Strong consistency on primary, eventual on read replicas.

Limits: 50,000 databases per account. 10 GB per database. 5 GB free.

OA relevance: High — every app's data layer.

R2 — Object storage ($0 egress) All

S3-compatible file storage. PDFs, images, videos, backups. Zero egress fees, forever. 10 TB of files/month on AWS S3 costs ~$900 in egress alone. R2: $0.

Pricing: $0.015/GB/mo storage. Free tier: 10 GB + 1M Class A ops.

OA relevance: High — client document vaults, media pipeline, backups.

KV — Key-value store All

Ultra-fast globally distributed storage for settings, feature flags, cached lookups. Eventually consistent (60-second propagation).

OA relevance: Medium — tenant branding, config, feature flags.

Durable Objects — Stateful actors with SQLite All

Each DO has a stable identity, its own private SQLite database, and WebSocket support. Single-threaded for consistency. WebSocket hibernation means millions of idle connections cost virtually nothing.

OA relevance: High — real-time collaboration, AI agent state, live dashboards.

Queues, Workflows, Containers All

Queues: Background job processing. User doesn't wait while heavy tasks run behind the scenes.

Workflows: Durable multi-step processes. If step 4 of 7 fails, it resumes from step 4. GA. Perfect for onboarding pipelines.

Containers: GA since June 2025. Run heavier software that doesn't fit the lightweight Workers model.

OA relevance: High — onboarding flows, report generation, document processing.

Workers for Platforms / Cloudflare for SaaS All

Workers for Platforms: Let your customers deploy isolated code on your platform. Each OA client can have custom business logic without a separate server fleet.

Cloudflare for SaaS: Serve many customers on their own branded domain with auto-SSL. portal.clientname.com just works.

OA relevance: High — the multi-tenant backbone for client dashboards and agent products.

Email Routing & Email Service All

Route incoming email for free. Send transactional email (receipts, password resets) from Workers. 3,000 emails/day included in $5/mo plan.

OA relevance: Medium — client notification emails, onboarding sequences.

AI, Agents & Browser Automation

The fastest-growing area. Directly relevant to OA's agentic product vision.

Workers AI — Serverless model inference All

50+ models (Llama, Mistral, Qwen, DeepSeek, Flux, BGE). One API call, no GPU to rent. Running in 200+ cities. 60-90% cheaper than OpenAI for tasks a smaller model handles well.

Pricing: 10,000 Neurons/day free. $0.011 per 1K Neurons paid.

OA relevance: High — AI features in every client product at low cost.

AI Gateway — Control plane for all AI providers All

Sits in front of Workers AI, OpenAI, Anthropic, Google, Groq. Caches responses, controls costs, retries failures, falls back between providers, logs everything. One dashboard for all AI spend.

OA relevance: High — cost control and provider flexibility for every AI feature.

Agents SDK & Project Think Sterling Gemini

Agents SDK: Framework for building autonomous AI agents. Each agent is a Durable Object with stable identity, SQLite memory, WebSocket connections, and built-in scheduling.

Project Think (@cloudflare/think): A higher-level, opinionated agent harness discovered by Gemini. Auto-wires message history persistence, stream resumption, and tool orchestration. Capability-based authorization means the agent starts with zero access and must be explicitly granted tools.

Capability Execution Ladder (Tiers 0-4):

TierWhat It Does
0 - WorkspaceLocal file operations backed by SQLite + R2
1 - Dynamic WorkerWrite and run code in a sandboxed V8 Isolate (zero network)
2 - npm PackagesDynamically install npm libraries, compiled on-the-fly
3 - Web AutomationSpin up headless Chrome to interact with websites
4 - SandboxFull secure OS container: clone repos, compile, run tests

OA relevance: High — the foundation for every AI agent product we sell.

AI Search (AutoRAG) — Chat with your documents All

Point it at an R2 bucket of documents. It automatically indexes them, keeps them in sync, and lets an AI answer questions grounded in your files. Supports external models (OpenAI, Anthropic).

OA relevance: High — "chat with your SOPs" for every client. The core of most business AI tools.

Vectorize — Semantic search database All

Stores the "meaning" of text as numbers so you can search by concept, not just keywords. Free tier: 30M queried dimensions/mo, 5M stored.

OA relevance: High — agent memory, document search, relationship intelligence.

Browser Run & Human-in-the-Loop Gemini ChatGPT

Runs headless Chrome on Cloudflare. The agent writes JavaScript CDP commands against the browser. Key features:

  • Live View: Real-time visual monitoring of what the agent sees
  • Session Recordings: Video + DOM logs for audit
  • Human-in-the-Loop: Agent pauses on captchas/logins, generates a secure link for a human to complete the task, then resumes

OA relevance: High — could replace our entire Mac Mini Puppeteer dependency. The human-in-the-loop pattern solves the Facebook login problem.

Remote MCP Servers Sterling

Cloudflare launched the industry's first remote MCP server (April 2025). MCP is the standard that lets AI agents securely connect to external tools and data. Cloudflare handles authentication (Auth0, Stytch, WorkOS) so users can safely grant an agent permission to act on their behalf.

OA relevance: High — agents that can interact with client's existing software.

WebMCP, x402 Payments, Code Sandboxes Gemini

WebMCP: An open standard for websites to declare their APIs for AI agents. Instead of scraping, agents discover structured endpoints.

x402: Machine-to-machine payment protocol (with Coinbase). Agents use micropayments to buy API access, compute, or data on the fly.

@cloudflare/codemode: Secure sandboxed environments where agents write and execute code safely. Zero network access in the sandbox.

OA relevance: Medium — forward-looking. WebMCP will matter when adoption grows. x402 is not immediately actionable.

Protect & Speed Up

The classic Cloudflare. Free or included on all plans.

CDN, DNS, DDoS, WAF, SSL/TLS, Bot Management All

All the foundational security and performance. CDN in 335+ cities. World's fastest public DNS (1.1.1.1). Unmetered DDoS protection (47.1M attacks mitigated in 2025, up 121% YoY). WAF blocks hacking attempts. Free SSL. Bot management and Turnstile (free invisible CAPTCHA).

OA relevance: High — every site and app gets this automatically.

Zero Trust / Access All

Put a login wall in front of any app. Control who gets in by email, Google login, or device. Free for up to 50 users. $7/user/mo after that.

Includes: Gateway (traffic filtering), Browser Isolation, CASB (SaaS security scanning), DLP (data leak prevention), WARP (device agent).

OA relevance: High — client and employee login for every dashboard.

Media, Analytics, Privacy

Stream, Images, Realtime, Analytics ChatGPT

Stream: Upload, encode, store, protect, and deliver live or recorded video. Training videos auto-adapt to each device.

Images: Store, resize, and optimize images on the fly.

Realtime: Low-latency audio/video for meetings and collaboration (SFU, TURN, RealtimeKit).

Web Analytics: Privacy-first traffic stats without cookies.

OA relevance: Medium — Stream for training content, Images for client dashboards.

Full capability count

ChatGPT cataloged 128 named capabilities across 12 categories. The Sterling Playbook covered ~30 in plain English with pricing. Gemini discovered 5 capabilities not in either (Project Think, Browser Run Human-in-the-Loop, WebMCP, x402, Code Sandboxes) plus 2 acquisitions (Human Native, Astro). This tab merges all three into one reference.

Client Proposal Engine

Building blocks for pitching Cloudflare-based solutions to OA clients. Pick a client need, get the architecture and pitch.

Pattern 1: Branded Client Dashboard

Client need: "I want to see my team's KPIs, download reports, and manage tasks in one place."

Architecture: Pages (frontend) + Workers (API) + D1 (data) + R2 (file storage) + Access (login)

Monthly cost: $5-50 depending on usage. If using Access beyond 50 users: +$7/user.

Pitch line: "A private, branded dashboard on your own domain with secure login, real-time KPIs, and document storage. No Google Sheets. No shared folders. Just your data, your way."

Pattern 2: AI Document Assistant

Client need: "My team wastes hours searching for SOPs and procedures."

Architecture: Pattern 1 + AI Search (pointed at R2 bucket of SOPs) + Workers AI or AI Gateway

Monthly cost: $20-100 depending on document volume and query frequency.

Pitch line: "Upload your SOPs and procedures. Your team asks questions in plain English and gets answers sourced from your own documents, with page references."

Pattern 3: Autonomous Agent

Client need: "I want an AI that monitors my email and CRM, drafts follow-ups, and schedules tasks."

Architecture: Agents SDK (or Think) + Durable Objects + AI Gateway + Remote MCP servers + Vectorize

Monthly cost: $50-200 depending on agent complexity and AI model usage.

Pitch line: "An AI agent that works 24/7, connected to your email and business tools. It drafts, schedules, and acts on your behalf, with your approval before any action."

Pattern 4: Multi-Tenant SaaS Product

Client need: OA builds a product once, sells it to many clients with isolation.

Architecture: Workers for Platforms + Cloudflare for SaaS (custom domains) + separate D1/KV/R2 per tenant

Monthly cost per tenant: $5-20 at the infrastructure level. Price to client: $100-500/mo with healthy margins.

Pitch line: "Each client gets their own private environment on their own domain. Your data never touches another client's data."

Pricing advantage to emphasize

Cloudflare does not charge egress fees. A client serving 10 TB of documents per month would pay ~$900 just in AWS bandwidth fees. On Cloudflare: $0. This is the single biggest cost differentiator in any proposal.

OA Internal Platform

How Outsource Access should use Cloudflare internally. Replace spreadsheets, secure internal tools, and power internal AI.

What We Already Have on Cloudflare

  • Brad's Command Center (brads-command-center.pages.dev) — Pages + D1 + Workers Functions
  • Sterling Pipeline Dashboard (sterling-workload.pages.dev) — Pages + static JSON
  • Client Dashboards (multiple Pages projects on OA Cloudflare) — Pages + D1 + Workers
  • Research Pages (research-for-brad.pages.dev) — Static Pages
  • Social Approval Dashboard (sterling-social-approval.pages.dev) — Pages + D1
  • Sterling Media CDN (sterling-media.pages.dev) — R2 bucket with public access

What Should Be Added (from Gemini's Blueprints)

  • Vectorize for semantic search across Command Center emails and contacts — "Who emailed me about the contract?" instead of keyword search
  • Email Worker for automatic email ingestion — currently polling Gmail API, could intercept at the edge
  • Durable Objects for real-time push updates to Command Center — currently using manual refresh/polling
  • Workflows for reliable multi-step onboarding processes — currently running sequences that can fail mid-step
  • AI Search pointed at OA's SOP documents in R2 — let the team ask questions about procedures

Internal Security Layer

Currently using custom auth tokens per dashboard. Gemini's recommendation: migrate to Cloudflare Access (Zero Trust) for proper identity-aware auth with Google login integration. Free for up to 50 users.

Benefits: one login across all internal tools, audit logs of who accessed what, device posture checks, no custom token management.

Estimated monthly cost for OA internal

Workers $5/mo base + D1 (included) + R2 ~$5-10/mo for document storage + Access $0 (under 50 internal users). Total: ~$10-15/mo for the entire internal platform. Compare to a single SaaS tool subscription.

Client Agentic Offering

AI agent products for OA's staffing clients. The revenue opportunity.

The Product Vision

OA already provides human VAs to clients. The next step: pair each VA with an AI agent that handles the repetitive work, letting the human VA focus on judgment calls and relationship work. The agent runs 24/7 on Cloudflare. The VA manages exceptions and approvals.

Agent Architecture (Synthesized from All Sources)

Inbound event (email, chat, webhook, schedule) triggers the agent. Here's the full stack:

LayerCloudflare ProductWhat It Does
Agent RuntimeThink / Agents SDK + Durable ObjectsPersistent identity, SQLite memory, WebSocket connections, scheduling
AI ModelsWorkers AI + AI GatewayMix cheap local models with frontier models (OpenAI, Anthropic). Cost control.
Knowledge BaseAI Search + Vectorize + R2Agent answers questions from client's own documents
External ToolsRemote MCP ServersAgent connects to email, CRM, calendar with proper auth
Browser ActionsBrowser RunAgent navigates websites lacking APIs. Human-in-the-loop for login/captcha.
Code Execution@cloudflare/codemodeAgent writes and runs code for data analysis, report generation
Client LoginAccess or Cloudflare for SaaSEach client gets their own branded login portal
Data IsolationSeparate D1 + R2 per tenantClient A's data never touches Client B's data

Security Model: Capability Execution Ladder

From Gemini's research. Each agent starts with zero permissions and must be explicitly granted capabilities:

  • Tier 0 (Workspace): Read/write local files only. Safe for any client.
  • Tier 1 (Dynamic Worker): Can write and run code, but no network access. Safe sandbox.
  • Tier 2 (npm Packages): Can install and use libraries. Still no network.
  • Tier 3 (Web Automation): Can browse websites. Needs explicit approval per domain.
  • Tier 4 (Full Sandbox): Full container. Clone repos, run tests. Maximum power, maximum oversight.

This maps perfectly to selling different agent tiers to different clients. A basic client gets Tier 0-1. A premium client gets Tier 0-3.

Pricing the Product

Your underlying Cloudflare cost per client at moderate scale:

ComponentMonthly Cost
Workers compute$5-20
D1 database$1-5
R2 document storage$2-10
AI model inference$20-100
Browser hours (if Tier 3)$5-50
Total infrastructure per client$33-185

Price to client: $200-500/mo per agent seat. That's 60-85% margin on infrastructure alone, before counting the VA management revenue.

The pitch to clients

"Your VA is great at judgment, relationships, and nuance. But they spend 60% of their time on repetitive data entry, email sorting, and status updates. Our AI agent handles that 60% around the clock, so your VA focuses on the 40% that actually grows your business. The agent is $X/mo. Your VA's impact doubles."

Consumer Products on Cloudflare

Brad's personal products and how Cloudflare powers them. Architecture blueprints from Gemini's research, cross-referenced with what we've already built.

Brad's Command Center

Current state: Pages + D1 + Workers Functions. Manual refresh/polling. Gmail API for email data.

Gemini's upgraded architecture:

  • Add Email Worker to intercept inbound email at the edge (replace Gmail API polling)
  • Add Vectorize for semantic search: "Who emailed me about the contract?" returns results by meaning
  • Add Durable Objects for real-time WebSocket push updates (no more manual refresh)

Stevens Family Planner

Current state: Pages + D1 + Workers. Family calendar, tasks, chores.

Gemini's upgraded architecture:

  • Add R2 for family document storage (insurance docs, medical PDFs, school forms)
  • Add Workers AI + Vectorize for "chat with your family documents" — "When does the home insurance expire?"
  • Add Durable Objects for real-time shared updates across family members' devices

DugoutReady

Current state: WordPress site on Cloudways (dugoutready.com).

Gemini's Cloudflare-native architecture:

  • D1 for game lineups and team profiles
  • Each active match gets a dedicated Durable Object
  • Coach moves a player → event via WebSocket to the DO → DO broadcasts to all connected devices instantly
  • Workers AI to suggest optimal lineups based on player stats and constraints

This is the architecture for the full product build whenever the WordPress site graduates to a real app.

Lanewise

Current state: brads-command-center.pages.dev (app) + lanewise.ai (landing page)

Potential additions: Agents SDK for an AI productivity coach that proactively suggests focus blocks based on calendar and energy patterns. Durable Objects for live collaboration if Lanewise ever goes multi-user.

Google vs Cloudflare: Complementary, Not Competing

Brad has both Google Workspace Enterprise (340 licenses + 120 Gemini Enterprise) and Cloudflare. They serve different roles.

The Complementary Model

DimensionGoogleCloudflare
Core strengthData, identity, device management, email, docs, calendarExecution, hosting, AI agent runtime, edge security
Best forSystem of record (who people are, what they wrote, when they meet)System of action (run code, serve apps, deploy agents)
AI approachGemini Enterprise: Agent Designer, connected data stores, Deep ResearchWorkers AI: serverless inference, Agents SDK, AI Gateway
AuthGoogle Identity (SSO, Workspace accounts)Access (consumes Google Identity, adds policy)
StorageGoogle Drive (collaboration, sharing, real-time editing)R2 (bulk files, $0 egress), D1 (structured data)
Pricing modelPer-user/mo ($12-25/user for Workspace, Gemini add-on)Usage-based ($5/mo base, pay for what you use)
HIPAAClear BAA across dozens of servicesBAA is Enterprise-only, developer products not publicly confirmed

How They Work Together at OA

  • Email: Gmail is the system of record. Cloudflare Email Workers can intercept and process inbound email at the edge before it hits Gmail.
  • Identity: Google Workspace is the identity provider. Cloudflare Access consumes Google login and adds per-app access policies.
  • Documents: Google Drive holds collaborative docs. R2 holds bulk files (PDFs, images, videos) at $0 egress.
  • AI: Gemini Enterprise for conversational search across Google data stores. Workers AI + AI Gateway for embedded AI features in apps at lower per-query cost.
  • Hosting: All web apps, dashboards, and client portals run on Cloudflare Pages + Workers. Google has no equivalent.

Where Google Wins (Don't Build on Cloudflare)

  • Regulated healthcare data: Google Cloud has clear, documented HIPAA coverage across dozens of services. Cloudflare's developer products are not publicly confirmed.
  • Connected data stores: Gemini Enterprise Agent Builder can connect directly to Gmail, Calendar, Drive, and Chat as data sources. Building this from scratch on Cloudflare would be significant work.
  • Collaborative editing: Google Docs, Sheets, Slides. Nothing comparable on Cloudflare.

Where Cloudflare Wins (Don't Build on Google)

  • Cost at scale: $0 egress, $5/mo compute base, usage-based pricing. Google Cloud charges egress and per-user fees.
  • Edge performance: 335+ cities, sub-millisecond cold starts. Google Cloud Functions have 100ms-2s cold starts.
  • Agent infrastructure: Agents SDK, Durable Objects, Browser Run are purpose-built for agents. Google's agent tooling is more focused on conversational search.
  • Multi-tenant SaaS: Workers for Platforms, Cloudflare for SaaS, custom domains with auto-SSL. Google has no equivalent.

The routing rule

Google = data, identity, collaboration. Use it as the source of truth for who people are, what they wrote, and when they meet. Cloudflare = execution, hosting, agents. Use it to build, deploy, and run everything that faces a browser or runs autonomously.

Account Architecture

Two Cloudflare accounts, clear separation rules, and multi-tenant design.

Two Accounts

AccountIDWhat Goes Here
Brad Personal (DEFAULT)f80e8ac3dae61d519e12427b58f8bd4dAll Sterling-built pages, dashboards, research, personal projects (Command Center, Lanewise, Family Planner), BST sites
OA Team79bf9918da35fa3f7d3a0bca83c0d0beClient dashboards, server.outsourceaccess.com, OA portal ecosystem, employee dashboards, anything under outsourceaccess.com

Routing Rules

  • Default: Everything goes to Brad's personal account unless explicitly directed to OA.
  • Keyword trigger: Brad says "OA Cloudflare" = deploy to OA account.
  • Auto-route to OA: Anything related to client dashboards, server.outsourceaccess.com, oa-portal-demo, oa-dash-directory, or any OA portal project.
  • API tokens: Personal = Keychain cloudflare-api-token. OA = Keychain cloudflare-oa-api-token.

Multi-Tenant Design for Client Products

When building a product that serves multiple OA clients:

  • Compute: Workers for Platforms gives each client their own isolated code environment
  • Database: Separate D1 databases per client (up to 50,000 per account)
  • Files: Separate R2 buckets or prefixed paths per client
  • Settings: Separate KV namespaces per client
  • Login: Cloudflare for SaaS gives each client their own domain with auto-SSL
  • Access control: Access policies scoped per client group

This architecture ensures client A's data, code, and files are completely isolated from client B's. No shared database tables, no shared file paths, no cross-tenant data leakage.

Security Baseline (from ChatGPT's Audit Checklist)

  • Identity: MFA for admins, individual accounts (no shared logins), automated joiner/leaver access changes
  • Authorization: Written role/permission matrix, authorization checked on every backend request, tenant isolation tested automatically
  • Files & data: R2 buckets private by default, short-lived signed URLs for downloads, retention/deletion rules documented
  • AI agents: Narrowest possible permissions per tool, secrets out of prompts, retrieved documents treated as untrusted, budgets and rate limits, human approval for irreversible actions
  • Operations: Separate dev/staging/prod, infrastructure as code, structured logs and alerts, test restoration procedures

HIPAA reminder

Both accounts are on commercial Cloudflare. If any client handles patient health data (PHI), get written BAA confirmation from Cloudflare Enterprise sales that Workers, D1, and R2 are covered before building. All three source reports agree on this.

Implementation Roadmap

A phased rollout plan synthesized from all three research sessions.

Phase 1: Foundation (Now — Next 30 Days)

Goal: Prove the core pattern works and establish internal confidence.

  • Migrate one high-value internal dashboard from Google Sheets to Pages + Workers + D1 + Access
  • Set up Cloudflare Access with Google login for all internal tools (free up to 50 users)
  • Establish R2 bucket for Sterling media and document storage
  • Document the build pattern as a reusable template

Benchmark to advance: The dashboard works, the team logs in with Google, and total cost is under $20/mo.

Status: Largely done — Command Center, client dashboards, and Sterling Media are already running. Access migration is the remaining item.

Phase 2: Add AI (60-90 Days)

Goal: Layer AI capabilities onto the existing infrastructure.

  • Add AI Search (AutoRAG) to Command Center, pointed at an R2 bucket of OA SOPs
  • Add Vectorize to Command Center for semantic email and contact search
  • Build one Agents SDK automation (daily briefing agent, email triage, or social media scheduler)
  • Set up AI Gateway for cost monitoring across all AI providers
  • Deploy one client-facing AI feature (document Q&A for a pilot client)

Benchmark to advance: A paying or pilot client uses an AI feature daily.

Phase 3: Productize (Quarter 2+)

Goal: Turn the repeatable architecture into a sellable SaaS product.

  • Set up Workers for Platforms for multi-tenant client isolation
  • Implement Cloudflare for SaaS so each client gets their own branded domain
  • Build the agentic product: VA + AI agent paired offering
  • Price per-seat with 60-85% margin on infrastructure
  • Build the Browser Run integration to replace Mac Mini Puppeteer dependency

Benchmark to advance: Multiple paying clients on the multi-tenant platform.

Phase 4: Scale (Quarter 3+)

Goal: Expand the agent offering and build competitive moats.

  • Expand to Tier 3-4 agent capabilities for premium clients (browser automation, code execution)
  • Integrate x402 micropayments if the protocol matures (agents autonomously purchasing data/APIs)
  • Build WebMCP integrations as third-party adoption grows
  • Evaluate Cloudflare Enterprise if any client needs HIPAA-grade PHI handling
  • Consider DugoutReady migration from WordPress to Cloudflare-native (Durable Objects for real-time)

Guardrails that change the plan

HIPAA: If any client needs real PHI handling, stop and get written BAA confirmation from Cloudflare Enterprise sales. Alternatively, keep PHI on Google Cloud or AWS and use Cloudflare only for the security front door.

Heavy compute: If an app needs a traditional always-on server or full Postgres, pair Cloudflare with Supabase (Postgres + auth) or Fly.io. Don't force it onto Workers.

Backups: Always keep D1 and R2 backups off-platform. Given the late-2025 and early-2026 outages, a Cloudflare outage should be an inconvenience, not a catastrophe.

Quick Reference: Product-to-Cloudflare Mapping

ProductCloudflare StackPhase
Internal dashboardsPages + Workers + D1 + Access1 (done)
Document Q&A / SOP searchAI Search + R2 + Workers AI2
Daily briefing agentAgents SDK + AI Gateway + Vectorize2
Client agentic productThink + Workers for Platforms + D1 + R23
DugoutReady (real-time)D1 + Durable Objects + Workers AI4
Family Planner AIR2 + Workers AI + Vectorize4