vupai 0.5.0

Voice UI for AI panes - push-to-talk voice control over tmux agent panes

Voice UI for AI panes: push-to-talk voice control for your tmux agent panes, on macOS, with on-device speech.

vupai (say "voo-pie") is a Voice UI for your AI panes.

Hold a key, speak, and what you say is typed into the right tmux pane: the one you're looking at, or an agent you call by name ("atlas, run the tests"). Speech-to-text runs on-device with NVIDIA Parakeet (via Apple MLX): no cloud, no API keys.

Built for a tmux-centric workflow where you keep several coding agents and shells open at once and want to drive them by voice without reaching for the mouse. New panes launch an agent by default (claude out of the box) and should work with other agentic coding tools (Codex, Gemini, …), though testing so far has focused on Claude Code.

▶️ Watch vupai drive tmux agents by voice

https://github.com/user-attachments/assets/fa2b4a62-87fe-45d3-9957-48e25852a005

Jump to: Requirements · Install · Set up · Usage · Voice commands · Supervision board · Activity ledger · Configuration · tmux tips · Uninstall

Why not plain tmux?

vupai runs on tmux: it doesn't replace it, it adds a voice layer on top. tmux already gives you panes, splits, and a way to keep many agents on screen. What it can't do is let you talk to them. That's the gap vupai fills.

With plain tmux	With vupai
Switch panes with <prefix>-arrow, then type	Hold a key and talk to the focused pane
Manually track which pane is which agent	Panes auto-name themselves; address them by name ("atlas, run the tests")
Re-type the same command in each pane	Broadcast by voice to every agent at once ("everyone, pull main")
Split / resize / re-layout with prefix chords	Voice commands: "create 3 panes", "focus atlas", "swap atlas and sage", "tile"
Read each pane yourself to see what agents are doing	Supervision board + "read atlas" speaks a one-line summary aloud
n/a	On-device speech (Parakeet via Apple MLX) - no cloud, no API keys

If you only have one shell open, you don't need vupai. It earns its keep when you are juggling several agents and want to drive them hands-on-keyboard-optional.

Because input is voice-first, it can also ease the typing load for anyone with RSI or hand-strain, though vupai isn't built or tested as a dedicated accessibility tool.

How it works

hold dictation key (Right-Option) → record (sox) → transcribe (Parakeet) → route → paste into a tmux pane → Enter

• Routing is hybrid. By default your speech goes to the focused pane. If you start with an agent's name, it goes there instead — even when it isn't focused. Say a number ("two, …") to hit a pane by its position in the current window.
• Injection is safe. vupai pastes your text and waits until it actually appears in the pane before pressing Enter — it never blindly submits.
• Local-first speech. Speech-to-text runs entirely on-device via Apple MLX (NVIDIA Parakeet): no cloud service, no API key, no account, so your voice and transcripts never leave your Mac. vupai itself makes no network calls beyond a one-time model download (~2 GB) on first use. The agents you drive, and the optional board summarizer, use whatever model you point them at, which may be a cloud service.

Requirements

[!IMPORTANT] vupai is macOS Apple-Silicon only: it depends on Apple MLX for on-device speech, plus two Homebrew binaries. It will not run on Linux or Intel Macs.

• macOS on Apple Silicon (M-series), macOS 13.5+ (developed on macOS 26).
• tmux and sox:

  brew install tmux sox
  

• Python ≥ 3.11 and uv, used to install the CLI:

  brew install uv          # or: curl -LsSf https://astral.sh/uv/install.sh | sh
  

Install

After the Homebrew step above, install the vupai CLI from PyPI in its own isolated environment with uv:

uv tool install vupai

This puts vupai on your PATH. The Parakeet model (~0.6B, ~2 GB) downloads automatically on first transcription.

To upgrade later: uv tool upgrade vupai.

[!NOTE] Prefer the bleeding edge? Install straight from git instead: uv tool install git+https://github.com/itsjrsa/vupai.

From source (development)

git clone [e-mail]:itsjrsa/vupai.git
cd vupai
uv sync            # creates .venv and installs everything (incl. the MLX runtime)

Run the CLI with uv run vupai … from the repo, or see the live-reload loop (vupai reload / vupai --reload) in AGENTS.md.

To install this checkout as a real vupai on your PATH (no PyPI), instead of uv run:

uv tool install .            # from the repo root; re-run after pulling changes

[!NOTE] The examples below use the bare vupai command (the installed tool). If you're running from a source checkout, prefix each one with uv run (e.g. uv run vupai setup).

Contributing? See AGENTS.md.

Set up (once)

After install, run the interactive bootstrap:

vupai setup

It handles everything first-run: checks the Homebrew tools, captures consent for the local transcript journal, lets you pick a mic and your push-to-talk key(s), downloads the speech model up front (so the first hotkey press doesn't stall on a silent fetch), then deep-links you to each macOS permission pane that still needs your terminal app enabled. Safe to re-run any time.

macOS permissions

vupai needs three permissions, granted to your terminal app (Ghostty / iTerm / Terminal / …) under System Settings → Privacy & Security: Accessibility, Input Monitoring, and Microphone. vupai setup deep-links you to each; to audit them on their own, run vupai doctor (it probes each and prints the exact System-Settings path for anything missing).

[!WARNING] macOS grants these to the terminal binary, not the script, so the hotkey and mic silently fail until you grant them. If voice seems dead, check this first (vupai doctor).

Usage

Start vupai inside a project, open a few agent panes, and drive them by voice. The push-to-talk daemon runs in the background, so you stay in tmux and just hold a key to talk. Launch (or re-attach to) a session with:

vupai                 # attach-or-create the session named after the cwd
vupai attach backend  # attach to "backend" (create it if absent)
vupai new backend     # create "backend" (error if it already exists)
vupai kill backend    # kill the "backend" session

[!NOTE] vupai runs on its own tmux server, so it never touches your existing tmux setup. The trade-off: its sessions don't show in a plain tmux ls; reach them with vupai attach. (Set tmux_socket = "" to share your default server.)

[!IMPORTANT] It's still tmux. The voice layer sits on top; every normal tmux binding keeps working. Detach with <prefix> d, split panes by hand (<prefix> % / "), switch with <prefix>-arrow, scroll/copy-mode, resize, your own custom keybindings: all unchanged. Use voice when it's faster, the prefix key when it's not. (Your ~/.tmux.conf is sourced too, with the few exceptions noted in tmux tips.)

Once attached, you talk to vupai with two push-to-talk keys:

Key	Config	Default	Hold and speak to…
Dictation key	hotkey	Right-Option	Type your words into the focused pane.
System key	command_hotkey	Right-Command	Run a voice command (below). The key is the signal, so there is no spoken control word; vupai acts on the panes instead of typing.

Both defaults are customizable: set hotkey / command_hotkey in the config (each takes a list, so you can bind several keys to one action).

vupai only listens for keyboard keys. To use a mouse button or other input as a push-to-talk key, remap it to a keyboard key (e.g. F13) with a tool like Karabiner-Elements or BetterTouchTool, then bind that key here.

Voice commands

Hold the system key and say any of these. Run vupai voice-commands for a cheat sheet tailored to your config.

Say	What happens
"create 3 panes"	Spin up N auto-named panes, tiled (up to 30; "create 2 shell panes" picks the program)
"focus atlas"	Focus the atlas pane (also "switch to / go to")
"swap atlas and sage"	Swap two named panes
"zoom atlas" / "unzoom"	Maximize a pane / restore the layout
"tile" / "layout …"	Re-layout the window (tiled, main-vertical, …)
"close atlas" / "kill atlas"	Close a pane (asks y/n by default)
"board"	Open the supervision board (one per session)
"open review"	Open a full-window diff review of the session (one per session)
"read atlas" / "read all"	Speak a pane's summary aloud (read board for a digest)
"clear atlas" / "clear all"	Send a slash command (/clear) to a pane or every agent
"everyone, pull main"	Broadcast the message to every named agent
"connect to box" / "ssh box"	SSH the focused pane into a configured host
"mute" / "unmute" / "stop"	Silence/restore talk-back, or cut off the current read
"louder" / "quieter"	Nudge readback volume (tts_volume, macOS say only)
"atlas, run the tests"	Not a command, so it falls through to name addressing

Commands

Run vupai --help for the full command list (and vupai <command> --help for a specific one). The everyday ones are in Usage above.

The push-to-talk daemon runs as a detached background process under your terminal app (not inside tmux — that's required for the global hotkey to work). It logs to ~/.config/vupai/daemon.log and survives detach/reattach.

Supervision board

When you have several agents running at once, you can't watch them all. The supervision board does it for you: vupai board (or just say "board") splits a dedicated pane (right, ~40%) that shows, per named agent pane, a one-line summary of its main conclusion or pending action — so a glance tells you who's done, who's stuck, and who needs you.

• Tool-agnostic. Works with any agentic CLI, not just Claude Code: pane activity is detected from terminal-output churn, and the summarizer is a swappable command (board_summarizer_cmd), not a fixed model. vupai appends the pane's scrollback tail as the command's last argument and takes its last stdout line as the summary, so any command that follows that contract works:

  To summarize with…	Set board_summarizer_cmd to
  Claude Haiku (default, streaming)	python -m vupai.claude_summarize --model claude-haiku-4-5
  Claude (plain, buffered)	claude -p --model claude-haiku-4-5
  Codex	codex exec
  Gemini	gemini -p
  Ollama (local/remote)	python scripts/ollama_summarize.py --host http://BOX:11434 --model qwen2.5:7b

  The model is whatever that command uses (e.g. Codex's own config/profile). If the command is missing or fails, the board falls back to a non-LLM last-line summary.

• Cheap by design. A pane is summarized only when it settles (finishes a burst of work), skipped when nothing changed, and throttled per pane (board_min_summary_interval).

• Speak it too. "read board" reads the digest aloud; "read atlas" reads a single pane.

One board per session. Close the pane to stop it.

Cross-pane activity ledger

When several agents share one working tree, they edit files unaware of each other and can clobber each other's uncommitted work. The activity ledger is a best-effort, pull-only awareness surface: a background poller records which pane last touched which file in each git tree, so you (or your agents) can spot overlaps before they become conflicts.

• What it records. Per git tree, in a .vupai/ directory at the tree root: activity.current.json (latest state per pane) plus activity.jsonl (history). Each entry names the pane, the files it touched, a coverage flag (exact when a tool edit-marker proves the pane wrote the file, git-delta when only scrollback ties it, or churn-only for an active pane no file could be pinned on), and any contended_with panes editing the same file. .vupai/ is auto-gitignored, so it never appears in git status.
• How it decides. git status provides what changed; each pane's scrollback provides which pane; their intersection is the attribution. A tool edit-marker (e.g. Claude's Update(<path>)) counts as proof a pane wrote that file and outranks a bare path mention, so a pane that merely discussed a file is not credited as its editor (and you don't get phantom conflicts from panes that just talked about it). It is post-write on a ~2s poll, so it reduces clobbering by surfacing overlaps; it does not prevent a sub-2-second race, and it never blocks or injects into a pane.
• Read it.
  • vupai activity shows the current ledger, grouped by tree.
  • vupai activity --stats reports contention and attribution rates (use these to judge whether it earns its keep in your workflow).
  • Say "activity" (or "who's editing") to hear the digest.

Let your agents use it

The ledger is pull-only by design, so nothing forces an agent to read it. If you want your agents to coordinate through it, tell them to: add a few lines to your project's AGENTS.md (or CLAUDE.md, or whatever instructions file your agent loads):

## Before editing a shared file
This repo may have several agents working in sibling panes at once. Before you
edit a file, read `.vupai/activity.current.json` at the repo root. If another
pane is listed as touching that file, or the file appears under a pane's
`contended_with`, stop and coordinate or pick different work instead of
overwriting it (the sibling's edits are uncommitted and you would clobber them).
A pane with coverage `churn-only` is active but its file is unknown: treat the
tree as contended and be cautious.

This is opt-in and best-effort: an agent consults the ledger only if its instructions tell it to, and even then it will not monitor the file continuously on its own. For a hard guarantee, give each agent its own git worktree (a planned opt-in) so panes physically cannot clobber one another.

Reviewing uncommitted changes (vupai review)

vupai review opens a live, full-window review of every uncommitted change across a session's panes, grouped by the pane that touched it. Run it inside a pane (it scopes to that session) or name one explicitly: vupai review backend. You can also just say "open review". It opens one review window per session, focusing the existing one rather than opening a second. It is a read-only, pull-only view: it runs git diff (the authoritative change set) and joins the activity ledger for attribution. It never stages, commits, or writes to a pane.

• Left: files grouped under the pane editing them, with +/- counts. Files touched by two or more panes are flagged ! and float to the top.
• Right: the selected file's diff, updating as you move. A file only one pane touched shows that pane's exact diff (its whole change is that agent's work). A file two or more panes edited shows the combined diff, flagged because it cannot be split per agent in a shared tree without worktree isolation.
• A trailing unattributed bucket lists changed files no pane claimed (including untracked files); names are never fabricated.
• It re-polls about every two seconds, so it tracks edits as they land.

Keys: up/down move, Enter (or o) opens the file in $EDITOR, space folds a pane group, p pauses live polling, r refreshes now, q quits.

Configuration

vupai reads ~/.config/vupai/config.toml. vupai setup writes it on first run, pre-filled with every key at its default and an inline comment explaining it, so the file itself is the reference. It's left untouched if one already exists, and vupai config --init tops it up with any keys a newer version added without disturbing your edits. Editing is optional; open the file to see them all.

Applying changes. The daemon reads its config once at startup, so a change takes effect only after it respawns. The interactive commands (vupai setup, vupai keys, vupai mic) apply their change automatically: if a daemon is running, they reload it for you. But edits you make by hand to config.toml or hosts.toml are not watched, so run vupai reload to pick them up (or vupai --reload to reload and attach in one step).

The keys most people touch:

Key	What it does
hotkey / command_hotkey	The dictation and system push-to-talk keys (pynput names; each a list, so you can bind several).
pane_command	Default program for voice-created panes (e.g. claude).
broadcast_word	Leading word that injects to every named agent (default everyone).
board_summarizer_cmd	Command that summarizes panes for the board and read (see Supervision board).
[programs] / [aliases] / [macros] / [slash_commands]	Spoken-token tables: program names, pane-name aliases, phrase macros, and slash verbs.

Remote machines (SSH)

The "ssh box" / "connect to box" voice command opens a new pane and SSHes into a host you name. Hosts live in a separate file, ~/.config/vupai/hosts.toml. Write a commented template with:

vupai hosts --init        # scaffold ~/.config/vupai/hosts.toml
vupai hosts               # list what's configured

Each host is one table; only host is required (SSH key auth must already work):

[hosts.box]
user = "me"               # optional; omit to use ~/.ssh/config defaults
host = "box.example.com"  # required: hostname/IP or an ssh-config Host alias
port = 22                 # optional
program = "claude"        # optional; omit to land in a plain login shell (default)

Say the table name ("ssh box") to connect. By default you land in a login shell, so you can cd into a project first; set program to auto-start an agent instead.

tmux tips

vupai sets the tmux options it needs at startup (ensure_up), so no config is required. A couple of optional settings just make the multi-agent flow nicer:

# ~/.tmux.conf (optional, pairs well with vupai)
set -g mouse on                       # click a pane to focus, scroll to read history
bind -T copy-mode-vi WheelUpPane   send -X scroll-up
bind -T copy-mode-vi WheelDownPane send -X scroll-down

[!WARNING] Do not enable extended-keys (CSI-u) in your tmux config:

set -s extended-keys on                     # breaks vupai
set -as terminal-features 'xterm*:extkeys'  # breaks vupai

It re-encodes Enter, so vupai's injected text never submits in Claude Code. vupai forces extended-keys off at startup, but a later tmux source-file (config reload) flips it back on and silently breaks submission. For the same reason, don't override pane-border-format / pane-border-status (clobbers the voice-name border) or rebind <prefix> + R (vupai uses it to rename a pane). These apply inside vupai's own session (tmux still sources your ~/.tmux.conf on vupai's dedicated server); your default tmux is untouched.

Roadmap

vupai is young and evolving. A few things on the horizon (in no particular order):

• Tighter pane-state and activity awareness so routing and the board react faster to what each agent is actually doing. (in progress)
• Worktree isolation (opt-in): give each agent its own git worktree so panes physically cannot clobber one another, which also makes cross-pane attribution exact. The hard-guarantee complement to the pull-only activity ledger and vupai review.
• Broader agent-CLI coverage: validate the flow end-to-end with Codex, opencode, Gemini, and other agentic tools (testing so far has centered on Claude Code).
• Smarter addressing: more forgiving name matching and disambiguation when several agents answer to similar names.
• More voice commands for everyday tmux moves, so less reaching for the prefix key.
• Linux support is a long shot (the speech stack is Apple-MLX-only today), but a pluggable transcription backend would open the door.

Ideas and contributions are welcome: open an issue or PR.

Uninstall

vupai down                       # stop the background daemon
vupai cleanup                    # revert any leftover settings on your default tmux server
uv tool uninstall vupai          # remove the CLI

That removes the program. To also delete what it created on disk:

rm -rf ~/.config/vupai           # config, hosts, daemon log, journal
rm -rf ~/.cache/huggingface/hub/models--mlx-community--parakeet-tdt-0.6b-v2   # the ~2 GB speech model

The Homebrew tools (tmux, sox) are general-purpose; remove them only if nothing else needs them (brew uninstall tmux sox). The macOS permissions were granted to your terminal app, not to vupai, so leave them unless you want to revoke them by hand under System Settings → Privacy & Security.

License

MIT. (Note: pynput is LGPL-3.0 and the Parakeet model weights are CC-BY-4.0; both are runtime dependencies, not part of this repo's code.)