Listening mode

Listening mode is SpeakEasy's explicit ASR input path. It turns one explicit utterance into the next turn of an exact Codex Desktop task, then narrates that task's real response through the native player.

The signed 0.2.19 release includes listening, voice lanes, the native player, and the paired-HTTPS Deck runtime. The Codex Desktop integration uses private, version-checked local IPC and is not a supported public Codex API.

Install SpeakEasy

The recommended path is to give Codex this bounded task:

Install SpeakEasy 0.2.19 on this Mac. Read https://speakeasy.arach.dev/agent.md and follow Path 3. Do not build from source or bypass Gatekeeper. Tell me which human-only steps remain.

Codex inspects the pinned installer before running it and reports the checksum, Gatekeeper, Developer ID, and exact-version checks. For a manual fallback:

1. Download SpeakEasy.dmg.
2. Drag SpeakEasy to Applications and open it.
3. Approve microphone and local-network access when macOS asks.
4. Open Settings → Deck and complete the three readiness checks.

The release is Developer ID signed, Apple-notarized, and self-contained. It supports Apple silicon Macs running macOS 14 or newer.

Start a conversation

1. Open the Codex Desktop task you want to control.
2. Open SpeakEasy's menu-bar popover.
3. Choose that task and lock it. SpeakEasy accepts the lock only after the Desktop window owning the exact task answers its validation handshake.
4. Press Control-Option-Space to start recording.
5. Speak one utterance, then press Control-Option-Space again to stop and send.
6. SpeakEasy transcribes locally, submits once to the locked task, waits for the correlated response, and plays it with the configured voice.

Starting another utterance while SpeakEasy is narrating stops its own playback first, preventing the app from transcribing itself.

Voice lanes

Nine persistent lanes make exact tasks addressable from anywhere on the Mac.

Assign a lane from the popover by choosing a recent Codex task. Every activation reopens and revalidates the exact task. An unassigned lane never guesses, and a failed validation leaves SpeakEasy unlocked rather than routing to another task.

Each lane can also carry its own narration voice and a compact spoken cue.

Privacy and routing

• Listening begins only from the explicit hotkey or control. There is no wake word, ambient listening, or automatic re-engagement.
• Capture stops on the second press or at a 120-second safety ceiling.
• Vox/Parakeet transcription runs locally on the Mac.
• Temporary microphone audio is deleted after transcription or cancellation.
• Raw input and transcripts are not added to SpeakEasy narration history.
• Cloud TTS is used only when you explicitly configure that provider.
• Task ID—not title, recency, directory, or window focus—is routing authority.

Codex compatibility boundary

SpeakEasy connects to Codex Desktop's current-user-only local IPC socket. It verifies protocol versions, socket ownership, task ownership, returned task and turn identifiers, and the response rollout associated with that exact turn.

If those checks fail, SpeakEasy stops with a visible error. It does not launch codex app-server, edit Codex state, paste through Accessibility, or fall back to a different task.

Because the seam is private, a future Codex Desktop update can require a SpeakEasy compatibility update. The long-term release should move to a supported extension or API when one becomes available while retaining the same exact-lock behavior.

Current non-goals

• ambient meeting transcription;
• wake words or voice activity auto-submit;
• inferred task selection from focus;
• a general-purpose voice assistant;
• a shadow Codex conversation;
• arbitrary tool or shell execution from speech;
• two-way Claude conversation.