AlphaFounder

AlphaFounder is an AI co-founder for pre-PMF solo founders — the people who have conviction and a problem but no teammate to think, build, and grind with them. Most early-stage founders either burn out alone or waste months waiting for the right co-founder; AlphaFounder compresses that loop into an always-on agent that onboards you over Discord, interviews you about your idea and constraints, builds a durable memory of your company, and then actively drives work: drafting investor updates, pressure-testing assumptions, producing artifacts, and holding you accountable to weekly momentum.

The system is a TypeScript monorepo deployed across Vercel and Fly.io. The marketing site and per-founder web app are Next.js 15 App Router with React 19, Tailwind v4, and Server Components. The backend is a Vercel Functions API on Fluid Compute, with Supabase Postgres as the system-of-record, Vercel Blob for artifacts, and Stripe for subscriptions plus credit metering. Each approved founder gets their own isolated OpenClaw agent — a long-running Fly.io Machine with a persistent workspace, HTTP-addressed, that runs Anthropic Claude (Opus 4.7 for deep reasoning, Sonnet 4.6 for fast turns) via OpenRouter so model choice stays a config flip. The Discord bot (discord.js, components v2, no embeds) is the primary founder interface and streams turns from the agent. The architecture is deliberately agent-first: states, timing, and pricing are hardcoded, but voice, judgment, and tool selection live in prompt artifacts (SOUL.md, modular SKILL.md files, and a memory layer) so improvement happens through skill edits and model upgrades, not more branching code.

Current progress: the full onboarding-to-active pipeline runs end-to-end in staging and production, with per-founder Fly provisioning, a real Stripe checkout flow, daily pulse syncs from each agent’s workspace back into Supabase metrics, and a deploy lane gated by remote synthetic canaries. A first wave of founders is being onboarded now; insights from every live run feed back into skill revisions rather than new code branches. The biggest risk we’ve addressed is prompt/agent drift — solved by keeping SOUL.md minimal and routing every behavioral fix to the specific skill that owns it, plus full-redo testing (fresh founder, fresh stack) on every change so regressions surface immediately.