flowviant

What it is

flowviant runs on your machines. A daemon sits on each one; the browser is a relay. You open a tab, pick a CLI — Claude Code, Codex, Antigravity — and the agent runs on your hardware, in your repo, with your credentials. Thinking streams back, keystrokes echo, a running terminal session can be adopted into a tab mid-flight.

The governing rule is that flowviant relays and never withholds. The only thing it knows that your agent doesn’t is whether a machine is reachable. Everything on screen is either something your agent said or something you’d have known by sitting at the keyboard. It has no reasons of its own to say “not now.”

What I chose

Deleting the server-side AI. flowviant used to ship its own project AI — a generation pipeline, a skills system, PRD drafting, a home-page wizard. It’s gone, deliberately and entirely. One brain reads the repo and does the work, and it’s the brain the user already pays for, running on their machine. The only model calls the server still makes are review quizzes, embeddings, and ticket triage, and a test pins that list so a fourth has to argue its way in.

Shipping to a daemon you don’t control. The web app and the daemon deploy on different clocks, and users update whenever they feel like it. Every capability sits behind a version floor: the daemon reports its version, and the server refuses a feature in words rather than failing strangely. Floors move independently of DAEMON_LATEST, so a feature can ship to the web before every machine has caught up.

Scale

A year of work, ~2,700 commits, three CLI runtimes, a published daemon at 0.48.x. It started as ASP.NET Core, moved to a TypeScript monorepo, and has been rewritten around sessions-as-tabs since.