BAMF Code: desktop vibe coding
A local Lovable meets Codex inside the BAMF desktop app. Prompt to live preview to visual edit to deploy, run against real repos with a worktree per task, diff review and PRs.
Build, ship, grow
Every app a creator builds is born inside their brand, their brain and their audience. That is the part Lovable and Codex cannot copy.
Brand-native
Apps inherit the brand kit and the shared design catalogue, so the first draft already looks professional.
Brain-native
The agent reads offers, voice, ICP and proof, so copy and pricing are real.
Ship equals distribute
Deploying to BAMF Sites drafts the launch post, wires lead capture and turns on analytics.
One mission control
Claude Code, Codex, Pi and BAMF Bots share one task list and one needs-you queue.
Less complexity
No second app or shell. Code reuses sessions, cards, BAMF Browser and Sites.
What it stands on
Most of the hard parts already existed. The missing piece was a native layer for files, processes and git.
Harness choice
We wrap the coding agents people already have, behind one adapter and one BAMF event model. Safety stays in deterministic code: a worktree per task, approvals for escalations, commands confined to trusted folders.
Claude Code
Stream-json CLI. Best approval controls, runs on the user's own plan.
Codex
app-server JSON-RPC. Cleanest embedding protocol, covers ChatGPT plans.
Pi
MIT, any provider or own key, for people with no plan.
Layout and journeys
One Code destination in the existing shell. A task is one worktree, one branch and one agent thread; several run in parallel.
Roadmap
Each slice in its own worktree, gated on build and tests, checked light and dark, then merged.
Session manager
One BAMF Bot triages every Claude and Codex session and only archives a session once its prompts, context, PR and worktree are accounted for.
Shipped
Desktop end-to-end test passed 7/7 with a real agent turn: register, worktree, edit, preview, commit.