The AI stack for indie hackers in 2026

The most confusing thing about building with AI in 2026 isn't the technology — it's the noise around it. Every week there's a new model, a new framework, a new primitive that promises to change everything. Most of it doesn't matter at indie scale. Here is the stack a solo builder actually needs.

This is not a list of every option. It's a recommendation. Everything here has been through real projects, not demos. The goal is a stack you can hold in your head, ship with in a weekend, and not regret in three months.

Pick one model and commit to it

The model layer feels like a weekly decision. It isn't. For most indie work, Claude Sonnet is the right call — it reads code well, follows multi-step instructions, and doesn't hallucinate APIs you have to look up. GPT-4o mini is the right call when you're optimizing for cost at scale: fast, cheap, and good enough for classification and extraction tasks. Gemini 2.5 Flash is the right call when the bill needs to be zero — the free tier is genuinely usable for prototyping. Pick one for the current project and don't switch mid-build.

Anthropic APIClaude Sonnet — the model that reads code like a senior engineer.anthropic.comOpenAI APIGPT-4o mini — fast, cheap, right for high-volume extraction and classification.platform.openai.comGoogle AI StudioGemini 2.5 Flash — capable, free tier, good for prototyping before you have a budget.aistudio.google.com

Your code assistant stack

The pairing that works: Cursor for navigating and editing the file tree, Claude Code for terminal-first, multi-step tasks — refactors, migrations, adding an entire feature with context across ten files. They complement each other. Don't try to replace either with the other. Windsurf is the alternative if Cursor's pricing bothers you — same model access, slightly different UX philosophy, worth a week of comparison before you commit.

CursorAI-native code editor — the fastest way to navigate and edit a codebase with LLM assistance.cursor.comClaude CodeTerminal-first AI coding agent — best for multi-file refactors and long-horizon tasks.anthropic.comWindsurfAI code editor from Codeium — solid Cursor alternative with a different collaboration model.codeium.com

Infrastructure that disappears

Vercel for frontend and API routes — Next.js on its home turf, zero-config deploys, built-in cron jobs, edge middleware. Supabase for everything else: Postgres, auth, file storage, realtime, Edge Functions — free to $25/month covers most indie projects end to end. Railway is the right fallback if you need more control: deploy anything via Dockerfile, no config overhead, pricing that doesn't punish you for running always-on servers.

VercelZero-config deployment for Next.js — cron, edge middleware, and analytics included.vercel.comSupabasePostgres + auth + storage + realtime as a managed service. Replaces your backend.supabase.comRailwayDeploy anything via Dockerfile with no platform config — the escape hatch when Vercel isn't enough.railway.app

Auth in a morning

Don't build auth from scratch. The attack surface is real and the maintenance tail is long. Clerk is the fastest path for most projects — prebuilt React components, handles email/password, OAuth, passkeys, and MFA out of the box, generous free tier up to 10,000 monthly active users. Better Auth is the open-source self-hosted alternative if you'd rather own the code and run it on your own database — same features, no vendor lock-in.

ClerkDrop-in auth components for React — email, OAuth, passkeys, and MFA, all handled.clerk.comBetter AuthOpen-source auth framework — own your session data, deploy on your own database.better-auth.com

The MCP servers worth wiring up now

Model Context Protocol is the emerging standard for giving AI agents structured access to tools and data sources. In 2026, three servers are worth wiring into Claude Code or any MCP-compatible client before reaching for anything more complex: the filesystem server (read and write local files by path), the fetch server (let the model make HTTP requests and read the response), and the Supabase server (query and mutate your database rows directly in conversation). These three together cover the majority of what an AI coding agent needs to be genuinely useful.

MCP FilesystemGives AI agents read/write access to local files — the foundation of any coding workflow.github.comMCP FetchLets an AI agent make HTTP requests and read the response — essential for API testing and scraping.github.comSupabase MCPQuery and mutate your Supabase database directly from an AI agent conversation.github.com

Skip LangChain. Use the Vercel AI SDK.

LangChain adds abstraction overhead that only pays off at team scale — chains, agents, callbacks, and retrieval components that take longer to understand than the underlying API calls they wrap. For a solo builder on Next.js, the Vercel AI SDK is the right choice: it handles streaming, tool calling, and multi-step agent loops with a small, readable API, and it's maintained by the same team as your deployment platform. For pure backend work without a UI, call the provider SDK directly — the overhead of an orchestration framework rarely pays for itself on a one-person project.

Vercel AI SDKStreaming, tool calling, and agent loops for Next.js — the right abstraction for solo AI projects.sdk.vercel.ai

The stack isn't what makes a product — the product makes the product. But the right stack gets out of your way and keeps the decision load low. Every tool here has one job, does it well, and won't become a liability when you need to move fast. Find them all on the Radar.