Danny

Not a chatbot. A chief of staff. Danny layers onto an OpenClaw agent and runs seven roles: dev, scheduler, todo-keeper, researcher, diary-keeper, odd-task runner, thinking partner. Each role is a behavior plus a file it reads and writes plus a trigger. The todos live in plain markdown. The diary is one file per day. Research lands in its own folder with the master-pattern citations at the bottom. The brain is files you own, not a database you rent.

Text Danny an app idea. It crafts the spec, then hands off. Claude Code writes every file. GitHub holds the private repo. Vercel builds and hosts. You get a live URL back. The handoff is async, so nothing times out: the tool returns a job id in milliseconds and a detached worker does the slow part. Danny itself writes zero code. It is the orchestrator, not the author.

Two keys. A build always stops once the repo is pushed, and waits. Nothing deploys until you reply 'ship it', and only from your Telegram. The cheap model runs the front of house with no shell, no browser, no cron, so an injected message structurally cannot act outward. Outbound actions route through a separate Pro sub-agent after your yes. Generated code never runs on the box; the only place it executes is Vercel's sandbox. The Google account is a throwaway, never the primary. The permissions file is root-owned and read-only to the agent.

Honest limits. Backends aren't auto-provisioned yet, so an app that needs a database won't get a real one until the Supabase wiring lands. Repo names can still collide. And the authoring bill is spent before the ship-it gate, so a build you reject still cost a few cents of model time. The whole system runs on a small VPS and one to five dollars a month of Gemini. The point was never scale. The point was one person, one Telegram thread, and a brain that ships.