Stacks

Build an AI agent

A TypeScript harness, a model API, a place to store state, and an editor that can keep up.

For a product that calls tools and keeps going, not a chatbot taped to a textarea. The framework is in your repo. The model is an API. The coding agent is how you build it.

Outcome An agent feature you own, with a model key, streaming UI, and a database for state.

1. 01

   Free

   Vercel AI SDK

   Harness

   TypeScript toolkit for streaming, tool calls, and UI. Start here if the agent lives in a web app.

   Score 8.5 · Streaming AI features in a TypeScript web app

2. 02

   Usage-based

   Anthropic API

   Model

   A strong default API for long coding and tool-use tasks. Swap in OpenAI or an OpenRouter key if you would rather not pick a single lab.

   Score 8.3 · Calling Claude from your own product

3. 03

   Free

   Mastra

   When workflows grow

   Move up to Mastra if the AI SDK starter turns into agents, workflows, and evals you want as a TypeScript framework.

   Score 8.0 · TypeScript agents you keep in your repo

4. 04

   Freemium

   Supabase

   Memory

   Threads, users, and tool results need a database. Postgres is enough. Do not invent a vector store on day one.

   Score 8.5 · Postgres plus auth for an indie app

5. 05

   Check site

   Claude Code

   Build faster

   Use a CLI agent to implement the harness against your repo once the shape is decided.

   Score 8.3 · Repo-sized edits from the terminal

6. 06

   Freemium

   Vercel

   Ship

   Host the Next.js app that exposes the agent. Serverless limits are real for long runs — check site if the agent must stay alive for minutes.

   Score 8.2 · Deploying a Next.js app with preview URLs