Built Project
VocsCode
One app, every agent
VocsCode started from a simple observation: every coding agent ships its own desktop app, and switching agents means switching interfaces — different threads, different review flows, different muscle memory. The agent matters; the shell shouldn’t.
VocsCode is one desktop app for all of them. Pick the harness per session — Claude, Codex, Pi, DeepSeek, Gemini CLI, or any OpenAI-compatible model — and everything works the same way through a unified interface. Switching agents mid-project changes nothing except the model doing the work.
What it does
Sessions, not conversations. Threads are grouped by project, each session gets an isolated git worktree, and diffs are reviewed per file — with revert and commit built in. Nothing gets lost when an experiment goes sideways.
Permission modes that agents actually respect. Ask, Accept edits, Plan, Auto, or Full access — selected per session, enforced identically across every harness. Dangerous commands always ask first.
Built for long days. Live model and effort switching, cost and context tracking, a real terminal in the side panel, slash commands, @file mentions, steer a running turn or queue the next one. API keys live in the OS keychain, never in the repository.
Status
Live and under active development. Setup is one command — clone, npm run setup, launch. A Doctor page detects installed runtimes and installs missing ones.
The source is public: github.com/vocsong/VocsCode.