The itch
Most agent demos take your question, go silent for a minute, and return a wall of text. I wanted to understand the other mode: an agent that searches the web while you watch and streams its answer token by token as information is retrieved, closer to watching a person research and write than to submitting a form. Lang Stream was the scaffold I built to learn how that fits together end to end.
How it works
LangGraph orchestrates the agent’s reasoning inside a FastAPI backend, which suits the job because streaming responses are native to it. The frontend is Next.js 15 with shadcn/ui. PostgreSQL persists state, Redis covers caching and message brokering, and the whole four-service stack comes up from a single docker-compose command with hot reloading on both sides. The API surface is deliberately tiny: POST a query to /api/v1/agent and read the stream.
Status
Archived. I built it over about two weeks in late 2025, got what I came for, and moved on. The code stays up under MIT as a working reference for wiring LangGraph, FastAPI and Next.js into one deployable unit. The honest post-mortem is in the case study.