06 — From Terminal to GUI, and the Business Model

GUI: admitting a mistake

Dax loves the terminal, but at ~8M monthly actives he thinks "too many users" are on it — a lot of them would have a better experience on a GUI, which the team understood very early 30:34.
They experimented with a web app and a desktop app (built with Tauri) but didn't treat the project seriously enough or think hard about the technical choices — "it ultimately was a mistake and we're moving back to Electron now" 30:56, 31:25.

Initially OpenCode (as an early Claude Code clone) stamped commits/PRs with its name, like Claude Code and GitHub Copilot do. When users asked to disable it, Dax reflected and killed it by default instead of adding an option 31:25, 32:30.
Reasoning: it felt like casino logic — trapping users with every little trick is the extreme of consumer products, and this hack is "too obvious; everyone knows why it's there" 32:12.

1. OpenCode Zen — inference service

Born to smooth onboarding: at launch, new Anthropic accounts couldn't even get rate limits high enough to use OpenCode, so they built a sign-up-once service giving access to all models 32:54.
It grew far beyond onboarding: hosting open-source models well is hard, so Zen became the place aggregating the best inference for them. Announced $50M run rate within ~5–6 months; margins on open-source models are decent 33:14, 33:37.

2. Enterprise control plane

"Extremely boring": a company with a thousand engineers can't just tell everyone to paste API keys — it needs providers, permissions, budget controls, and rate limits centrally managed. Currently enterprise-deployed, going public soon; open source, but most pay for the hosted version 34:00.
Timing tailwind: companies are finally scrutinizing LLM spend just as open-source models became competitive at ~10× cheaper — so the control plane naturally funnels customers into Zen inference. If inference dominates, they may stop charging for the control plane entirely 34:25, 34:42.