Tandem

Inspiration

I've launched 10 SaaS products at Boxford Partners -- TitleWise, StrideTC, ReviewSniper, and more. Every single one started with the exact same first sprint: provision a Cloudflare Worker, set up Stripe, configure Postmark, write a privacy policy, generate a sitemap, draft a service agreement. It took days every time. Tandem was born from the frustration of repeating that sprint over and over. The idea is simple: what if you could describe a business and have a team of AI agents do that entire first sprint for you, in seconds?

What it does

Tandem is an AI technical co-founder. You have a natural conversation about your business idea -- target customer, pricing, design preferences, features -- and when you approve the plan, 8 autonomous agents execute in parallel:

Planner generates a creative product name, features, FAQ, and design direction via Claude
Infrastructure provisions a Cloudflare Worker with a full landing page
Payments creates a Stripe customer, product, price, and hosted checkout session
Comms sends a branded HTML welcome email via Postmark
Legal drafts a tailored privacy policy and terms of service
SEO generates robots.txt, sitemap.xml, and AI-crawler-friendly configuration
Research analyzes the competitive landscape and identifies positioning, risks, and quick wins
Contracts drafts a service agreement with scope, payment terms, and liability clauses

Then the Deployer rebuilds the Worker with everything baked in: checkout URL in the CTA, legal docs at /privacy and /terms, service agreement at /agreement, sitemap at /sitemap.xml, and an llms.txt for AI agent discoverability.

Every spawned site is live on the internet. Every Stripe checkout collects real money. Every email lands in a real inbox.

How we built it

Tandem runs on Cloudflare Workers (Hono) serving a React SPA with Tailwind 4. Project state lives in Cloudflare D1. The conversational co-founder uses Anthropic Claude Sonnet 5 with adaptive thinking to gather requirements before building anything.

Brainbase orchestrates the agent team -- 9 agents registered with a 12-edge directed acyclic graph that maps the execution topology. The Planner fans out to 7 specialist agents in parallel, then the Deployer waits for all dependencies before the final build. The orchestration graph is exposed at /api/brainbase and agent IDs are visible in the UI during execution.

Each agent calls real APIs:

Cloudflare API for Worker provisioning and subdomain activation Each agent calls real APIs:
- Cloudflare API for Worker provisioning and subdomain activation
- Stripe SDK for customer/product/price/checkout creation
- Postmark API for branded HTML email delivery
- Claude for legal documents, competitive research, service agreements, and FAQ generation

The landing pages adapt to design preferences captured during the conversation -- light or dark theme, custom accent colors, gradient mesh backgrounds. Every page ships with FAQPage JSON-LD schema and an llms.txt endpoint for AI discoverability.

The UI features a split-panel layout (chat left, output right on desktop, toggle on mobile), a live launch checklist with progress bar, a real-time technical log showing what each agent is doing, and light/dark mode. Slack notifications fire at every agent action.

Challenges we faced

The biggest challenge was making Claude API calls work within Cloudflare Worker execution limits. The hackathon API key required a workspace ID header and adaptive thinking mode -- both undocumented quirks we had to debug live. Our first attempt at fully Claude-generated landing pages (having Claude write the entire HTML) timed out the Worker. We solved this with a hybrid approach: Claude generates the content (names, features, FAQ, legal docs), and a design-aware template renders it.

The second challenge was getting both panels to scroll independently in Tailwind 4. The flex height chain didn't resolve correctly across multiple nested containers. After five failed attempts with different flex/overflow strategies, we abandoned Tailwind utilities for layout-critical properties and used position: fixed with explicit dimensions -- the only approach that worked reliably.

What we learned

Agent orchestration is more about data flow design than AI prompting. The hard part isn't making Claude generate good content -- it's making 8 agents coordinate their outputs so the Deployer has everything it needs. The Brainbase DAG made this explicit and debuggable.

We also learned that the "co-founder" metaphor changes everything about the UX. When we treated it as a chatbot that deploys templates, it felt thin. When we added clarifying questions, design preferences, post-deploy recommendations, and next-step roadmapping, it felt like working with someone who actually cares about your business.

What's next

Persistent memory via Mitosis Labs so the co-founder remembers every decision across sessions
Workers for Platforms for multi-tenant customer Cloudflare accounts
Iterative buildout -- after the first deploy, Tandem walks you through security, auth, integrations, AEO optimization, and compliance
Document signing via Documenso integration for the Contracts agent
Intake form for post-deploy content enrichment and site redeployment