The indie agent builder stack
A pragmatic TypeScript stack for solo or two-person teams shipping AI agents in 2026. Optimized for fast iteration, low ops overhead, and graduation paths when the product hits scale.
- 01CLAI Coding Agentsai-coding-agent
Claude Code
Terminal-native autonomous coding with Skills for repeatable workflows.
- 02CLAI Modelsllm
Claude Opus 4.7
Sharpest frontier model for agentic loops and long-context work.
- 03MAAgent Frameworksagent-framework
Mastra
TypeScript-native with workflows, evals, and RAG built in.
- 04ASWeb Frameworksweb-framework
Astro
Static-first marketing site plus hybrid pages for the app surface.
- 05VEHostinghosting
Vercel
Best-in-class DX, generous free tier, preview deploys per PR.
- 06SUDatabasesdatabase-auth
Supabase
Postgres plus auth, storage, realtime, and edge functions in one box.
Why this combo
This is the "ship in a month, scale to ten thousand users without a migration" stack. TypeScript end-to-end; one cloud provider for the data plane; one for hosting; one provider for the model; one framework each for the marketing site and the agent layer. Every layer has a clear graduation path if the product warrants it.
This is the stack we’d hand a solo developer or two-person team shipping their first agent-driven product in 2026. It is intentionally boring on the foundation and intentionally sharp on the AI surface.
Why this combination works
One TypeScript codebase. Mastra is TS-native, Astro is TS-native, Supabase has clean TS clients. No language boundary, no API contract maintenance, no JSON-shape drift between services.
One cloud per concern. Vercel handles hosting and edge compute. Supabase handles data, auth, and storage. Anthropic handles inference. Each provider has a generous free tier; the combined free-tier ceiling covers more usage than most v1 products will see in their first month.
Clear graduation paths. When you outgrow Supabase, you move to Neon plus separate auth (Clerk) without rewriting your application layer. When you outgrow Vercel, you move to Cloudflare or self-host. When you outgrow Mastra’s surface area, you’ll likely have enough domain knowledge to know what you actually need.
What this doesn’t cover
Observability, evals, and cost monitoring are not in this v1 cut. At indie scale, console logs plus Mastra’s built-in evals plus Vercel’s analytics is enough. At team scale, you’ll want Langfuse or Helicone in the loop.