v0.1.1

Open Room 0.1.1

Fixes from the first days of real use, all from field reports against 0.1.0. Same footing as before: unsigned builds
(see the 0.1.0 notes for the SmartScreen and Gatekeeper steps), everything local, every agent bills your own Claude
Code login. The macOS build remains untested — a report either way is still the most useful thing a Mac user can send.

Voice input

• Long spoken prompts are no longer cut off. Capture had a hard 30-second limit that ended listening mid-sentence and
  silently sent the truncated text. The limit is now a five-minute stuck-microphone failsafe, and if it ever fires the
  listening pill says so.
• Pausing to think no longer ends a capture. The silence hang went from 1.5 s to 2.5 s; pressing the key again still
  ends it instantly.
• Starting to talk right after pressing the key no longer deafens the capture. The noise floor is sampled robustly and
  can only ever adapt downward, so a floor set during your own speech repairs itself at the first pause.
• Push-to-talk and wake words are independent. Each has its own switch; neither requires the other. Hands-free-only
  use works.
• The tray's voice item is a true off switch: one click turns off both push-to-talk and wake words and closes the
  microphone.
• Enabling voice offers the speech model download in place. The switches are no longer disabled until Whisper is
  installed — flip one and you are offered the one-time download right there; declining leaves it off.

Interface

• Dialogs no longer close on a stray click outside them. The agent editor and settings dialogs pulse instead, and
  close only on Esc, the X, Cancel or Save. Half-edited agent config is no longer lost to a mis-click.
• The context meter warns relative to when the session will actually compact, read from the SDK, rather than at fixed
  70% and 90% marks.

v0.1.0

Open Room 0.1.0

First release. Open Room is a desktop app for running multiple named Claude Code agents at
once — each with its own model, tools, voice and persistent role — addressed by typing,
push-to-talk, or an opt-in wake word ("hey Atlas…"). Agents work concurrently, report back
by native notification or text-to-speech, and their conversations persist and resume across
launches.

Before you download: Open Room ships no credentials and runs no cloud of its own. Every
agent runs through the Claude Code login already on your machine, and everything an agent
does bills your own Claude Code account — exactly as if you had typed it into the CLI
yourself. You need Claude Code (https://www.anthropic.com/claude-code) installed and signed
in (npm install -g @anthropic-ai/claude-code, then claude to sign in). Everything is local:
no telemetry, no analytics, no cloud speech — speech recognition and synthesis run on your
machine.

Installing

• Windows: run open-room-0.1.0-setup.exe.
• macOS: open open-room-0.1.0.dmg and drag Open Room to Applications.

The builds are not code-signed, so both operating systems will object once:

• Windows: SmartScreen shows "Windows protected your PC". Click More info, then Run anyway.
• macOS: Gatekeeper says the app "cannot be opened because the developer cannot be
  verified". Right-click the app and choose Open, then Open again in the dialog — or run
  xattr -d com.apple.quarantine "/Applications/Open Room.app".

Nothing is bundled beyond the app: no voices, no speech models. System text-to-speech works
immediately; anything neural (Whisper for voice input, Kokoro voices) downloads on demand,
with size and licence shown before you agree. Voice input ships off; the wake word is a
second, separate opt-in.

⚠️ The macOS build is untested

The DMG is built and configured (entitlements, microphone usage description) but no one has
ever launched Open Room on a Mac. If you try it, please open an issue
(https://github.com/TheLinc/open-room/issues) with what happened — even "it does not start"
is genuinely useful, and "it works" doubly so. The Windows build has been installed,
launched and exercised end to end.

Known gaps in 0.1.0

• Hold-to-talk does not exist yet — voice capture is toggle-to-talk (press to start, press
  again or go silent to send, Esc to discard).
• Builds are unsigned on both platforms.
• The always-on listening CPU figure (~8% of one core) was measured on a desktop, not the
  older laptop it was budgeted for.

The full list, including what is deferred by design, is in CONTRIBUTING.md
(https://github.com/TheLinc/open-room/blob/master/CONTRIBUTING.md).