ThiagoEMatumoto/pitwall

A desktop cockpit for running and supervising many Claude Code sessions across your projects and repositories.
English · Português (BR)
Watch the intro video
Claude Code is a terminal-first agent. Pitwall wraps it in a desktop app so you can drive several agents at once — each in its own project and repo — without juggling terminal tabs. It keeps a local SQLite database of your projects, sessions, plans and repositories, renders each session as either a live terminal or a structured chat, and connects related repositories so a session in one repo can delegate work to a session in another.
The name comes from the Formula 1 pit wall: you are not driving the car, you are on the radio with several drivers at once.
> Status: actively developed, versioned releases for macOS, Windows and Linux. Some features are marked experimental below.
## Voice: dictate a prompt, hear the answer
The point of the voice mode is what it does not do: dictation never sends anything. The transcript lands in the composer for you to read, edit and send yourself.
Dictation. The mic button records in the renderer (getUserMedia + MediaRecorder, webm/opus) and posts the audio to an OpenAI-compatible /v1/audio/transcriptions endpoint. The endpoint, key, model, language and a minimum-duration threshold come from ~/.config/voz/voz.env — nothing is hardcoded to a vendor. Recordings shorter than the threshold are discarded instead of transcribed.
Long dictation is condensed, not rewritten. A speech transcript carries hesitations, restarts and mangled English terms ("abrir com request" → "abrir pull request"). Long dictations go through claude -p with haiku (tools disabled — the spoken text must never trigger an action) before reaching the composer. Two guards: the pass is fail-open, so a failure returns the raw dictation instead of losing it, and a result more than 1.2× the original is treated as the model explaining rather than condensing and gets discarded.
Spoken turn summaries. When a session crosses the working → waiting/idle edge, Pitwall reads the last assistant turn from the transcript and turns it into 2-3 spoken sentences (also haiku), streamed through ElevenLabs. It only fires for sessions where you turned it on, one by one — a per-session opt-in, because the cost is per turn. A permission prompt is not the end of a turn and does not summarize. There is also a button to summarize on demand, which ignores the opt-in gate because you asked for it explicitly.
## Diagrams: an Excalidraw canvas the agent can edit
Diagrams are a real Excalidraw canvas, editable by hand. The agent writes to it through MCP tools instead of producing an image.
diagramcreate takes a skeleton: semantic nodes and arrows, ids instead of coordinates. Omit x/y and the converter lays it out; arrows reference node ids via start/end and carry a label. Raw Excalidraw scenes are accepted too, but the skeleton is the preferred input.
diagrampatch is the reason this works. It applies incremental add | update | delete ops addressing elements by skeleton id on top of the current scene, so the box you dragged and the color you changed survive the agent's next edit. diagramupdate exists for a deliberate full redraw and says out loud that it discards your refinements. Both record a version snapshot with a changelog line.