Open source · MIT · runs on your Mac

Talk to all your coding-agent sessions by voice.

Ask what each one did, switch by saying its name, and answer permission prompts out loud, from across the room, under rules enforced in code. It works with Claude Code, OpenCode and Codex, with no server and no account: you bring the model keys, or run every model on your Mac.

claude plugin marketplace add https://github.com/ballerine-io/pointe.git && claude plugin install pointe@pointe

Then run /pointe:setup in a new Claude Code session. Requirements

The Pointe iPhone app during a call: billing/main needs approval for git push --force-with-lease origin retry-fix, with Deny and Approve.

Demo video coming before launch

Until the demo is recorded, this is a still of the iPhone app from the repository's screenshots.

One conversation

The sessions do the work. Pointe listens, reports and routes.

Pointe is one voice that knows every session you join to it. Each session keeps its own context and keeps working. Background sessions stay quiet except for "finished" or "needs approval". A "by the way" question is answered from a read-only copy, so the live session is never interrupted.

An example call

You
What is the payments session doing?
Pointe
It finished the retry fix. Tests pass. It wants to know if it should open the PR.
You
Yes. And tell the docs session to update the changelog.
Pointe
Docs is on it.
You
By the way, why did payments drop the queue idea?
Pointe
Duplicate charges during failover. It decided that an hour ago.

How it works

A daemon on your Mac, between your voice and your sessions.

A Python daemon built on Pipecat runs the call. The voice reaches it from a Chrome tab on the Mac or from the iPhone app. Each session you join gets an adapter that follows its record, sends your requests to it, and holds its permission prompts.

  1. You

    Chrome page on the Mac

    Loopback only. Live transcript, the exact text of each pending request, Approve and Deny.

    iPhone app

    Paired once by QR code over Tailscale, a VPN or Wi-Fi, with a pinned TLS certificate. Approve needs Face ID.

  2. The daemon on your Mac, the hub

    Earsspeech to text Echo guarddrops the bot's own voice Braindecides through tools; each reply is checked in code Mouthtext to speech

    Approvals

    The read-out allowlist, the bound yes, the tap, the optional voice lock (a local model).

    Gate

    Each send to a session is checked against what you just said.

    Reader

    Session notes, reports, briefings, conflict checks, meeting notes.

  3. Your sessions

    Claude Code

    Joined with /pointe:voice. Approvals through its PermissionRequest hook.

    OpenCode

    Version 2. Joined with pointe join opencode.

    Codex CLI

    Started as codex --remote unix://, joined with pointe join codex.

    Other machines

    A node on another Mac or a Linux host runs the same adapters, with no audio and no keys, over a pinned, authenticated link.

Each role (ears, brain, reader, checker, mouth) is set separately. A role uses the vendor you chose, under your key, or a model on your Mac. Detail: how-it-works.md.

Features

What it does.

One voice over many sessions

Say a session's name to talk to it or switch to it. Ask what is going on across all of them.

Session notes

A reading model keeps each session's goal, state, decisions, what was ruled out and what is open, and turns each final message into what you need to hear.

Side questions

"By the way" questions go to a read-only copy of the session, so the live one keeps working and gets no new message.

Live transcript

The page shows the conversation as it happens and lets you type or correct what was heard.

iPhone app

The same call and controls on your phone. Approve needs Face ID.

Several harnesses and machines

Claude Code, OpenCode and Codex side by side in one call, each session with its own approvals, on this Mac or on other machines.

Each feature page says how to use it, the rules, what was measured, and the limits.

A call over three sessions: a report from billing, a corrected request, a typed side question and its answer.
A call over three sessions
The sessions sheet: river-lantern on Codex needs you, billing/main on Claude is working, docs on OpenCode is idle.
Sessions by agent
Fix it: what was heard, tell building to bump the timeout to 30 seconds, and an editable field for what you meant.
Fix what was heard

Safety model

Spoken approvals, under rules in code.

Words you speak can make a session do work, so the rules are in code with tests, not in a prompt.

What can approve

  • Your spoken yes, only for a request the daemon read out itself, from what will run, not from the model's description. It must be the last thing the bot said, and a plain yes, begun after the read-out ended, within 120 seconds, on the active session.
  • A tap on the exact request, on the page or in the iPhone app. A request the daemon cannot read out whole (unknown programs or flags, $, globs, subshells, several lines, sudo, more than 25 words) is approved only this way.
  • The terminal, as always. Your Claude Code deny rules still apply.

What never approves

  • "Okay, tell me more", "yes, and run the linter", silence, or a yes to any other question.
  • An echo of the bot's own voice.
  • Typed words, in the call or as a correction.
  • A voice match: the voice lock only narrows, and a match is never consent.
  • A mode: drive, whisper, meeting, briefing and challenge modes can only narrow what a word may do.

Always

  • A no from any voice denies a pending request, in every mode.
  • Secrets are never spoken. A key, a token or password= is never read out or put in a briefing, and is masked on screen.
  • Voice is not authentication. Anyone your microphone hears can say no, and, with the voice lock off, yes. Do not use voice approvals in a room you do not trust.

About 13%

of shell commands in our own Claude Code history can be read out, so most commands need a tap. That is deliberate.

Set "approvals": "terminal" in ~/.claude-voice/config.json to keep every prompt in the terminal. The daemon listens on loopback only, every route needs a secret, and Origin must match Host. The threat model, including what Pointe does not defend against: SECURITY.md.

Privacy

What leaves your Mac, and where it goes.

Pointe has no server and no telemetry, and the call page loads nothing from outside your Mac. What leaves the Mac goes to the vendors you chose, under your own keys.

WhatGoes to
Your speech, plus hints: your dictionary, session and folder names, names from session notesThe ears vendor: Soniox, Google Gemini, OpenAI or ElevenLabs
Session content: transcripts, notes, reports, briefings, conflict checksThe brain and reader vendors: Anthropic, Google Gemini or OpenAI
What the bot says, which can quote session contentThe mouth vendor: ElevenLabs, Google Gemini, Soniox or OpenAI
What you said, each session's latest lines, goal, folder path and up to 40 file namesThe gate vendor
In a meeting you started: the room's audio, and transcript windowsSoniox, and the reader vendor

pointe local voice: voice stays on your Mac

Whisper hears and Kokoro speaks on the Mac. No audio goes to any vendor, and meeting mode is off. Your words reach the brain, reader and gate vendors as text.

pointe local: fully local

Whisper, Kokoro and Gemma 4 26B A4B on the Mac, and no key is read. Needs an Apple-silicon Mac with 32 GB of memory and about 19 GB of disk. Slower and less accurate, by measure: local.md.

Always on the Mac

Your voiceprint, the call journal, session notes, the dictionary, briefings and the files shown on screen. Keys come from the environment or the macOS Keychain and reach the daemon only.

With push to talk on, no audio leaves the Mac unless you hold the talk control (plus 0.4 s after release). Full list: SECURITY.md. This page has no analytics and loads nothing from other hosts.

Measured latency

From your last word to the first sound.

1.86 s

median

24 turns on a real call

2.21 s

p90

slowest turn 4.19 s

3.17 s

median before 1.14.0

p90 4.27 s, 12 turns

How it was measured

tests/e2e/latency.py asks twelve plain questions on a real call, with a synthesized voice as the user, and stamps each turn stage by stage. Real vendors: Claude Sonnet 5.5 as the brain, Soniox ears, ElevenLabs mouth. One 18-core Mac that other work kept busy (load 4.2 rising to 7.6). It is reproducible with your own keys; it calls paid APIs.

Source: conversation-flow.md §7.

Stages, median and p90 in seconds. They overlap: the model is asked before the turn is taken.
StageMedianp90
Ears: last word to final transcript0.460.52
The turn is taken, after the last word1.081.10
Model: request to first token0.740.82
Mouth: request to first audio0.220.33
Client: playing to audible0.070.11
Last word to first sound1.862.21

Install

One command, then a guided setup.

You need macOS, Claude Code 2.1.224 or newer, Chrome, headphones, and one model key, or an Apple-silicon Mac with 32 GB of memory for the fully local mode.

  1. Install the plugin

    claude plugin marketplace add https://github.com/ballerine-io/pointe.git && claude plugin install pointe@pointe
  2. Run setup in a new Claude Code session

    /pointe:setup

    It checks your Mac, installs uv if it is missing, helps you choose vendors (with measured quality and cost), stores each key in the macOS Keychain through a Mac window (never in the chat), starts your first call, and can pair your phone.

  3. Join each session you want to reach

    /pointe:voice

    The first run downloads about 180 MB. A tab opens: allow the microphone, press Connect, and talk. OpenCode and Codex sessions join with pointe join opencode or pointe join codex.

You pay the vendors directly. A busy hour costs about $1 on the best measured setup (Anthropic, Soniox, ElevenLabs and Gemini keys), by arithmetic from measured use and list prices, not a metered bill. Other setups: benchmarks.md. More install options, including the Claude desktop app: install.md.

Compared to

Other ways to talk to a coding agent.

Pointe is not the first tool to read permission requests aloud. These are the closest, described from their public documentation, read on 2026-11-19.

Happy

Open source (MIT). Voice docs · repository

  • Its voice agent reads permission requests out loud and allows or denies one only when you say so, unless you tell it to accept the ones that follow.
  • It starts with a list of your active sessions, follows you when you switch, and can send a message to a session in the background.
  • Voice runs on ElevenLabs: free for 20 minutes every 30 days, 5 hours with a subscription, or unmetered with your own ElevenLabs agent.
  • Apps on the App Store and Google Play, and on the web. Sessions start with happy claude or happy codex; a relay server syncs your devices, end-to-end encrypted.

Where Happy is ahead: store apps for iPhone and Android, more platforms, and many more users (over 24,000 GitHub stars).

Where Pointe differs: the approval rules are code with tests, and a request the daemon cannot read out word for word needs a tap. There is no relay server, account or subscription: the daemon runs on your Mac and you pay each vendor directly, or run every model locally. Claude Code and OpenCode sessions you already run join as they are; Codex must be started with --remote. Each session gets notes, and side questions go to a read-only copy.

Claude Code /voice

Built into Claude Code. Voice dictation docs

  • Dictation: hold or tap a key, and your speech is transcribed into the prompt input. In agent view it can dictate to a background session.
  • Audio streams to Anthropic for transcription. It needs a Claude.ai sign-in, and transcription does not count toward your usage limits.
  • Runs on macOS, Linux and Windows, in 20 dictation languages.

Where /voice is ahead: nothing to install, no model keys, and more platforms.

Where Pointe differs: it talks back. One spoken conversation covers several sessions and harnesses, reports what each did, and reads approvals out for a yes or a tap.

Limits

What it does not do yet.

  • The Mac is the hub. The call runs on macOS with Chrome. Other Macs and Linux hosts join as nodes, with no audio. On Linux the daemon starts, but no real call has been made there. Windows is not supported.
  • iPhone app: build it in Xcode for now. It is not in the App Store, and not on TestFlight yet; a TestFlight build is planned. With a free Apple team the build opens for 7 days. Without the app, the page works in a phone browser. phone.md
  • Synthetic evidence. The benchmarks, the voice lock thresholds and the whisper numbers use synthetic voices. No bench on recorded people has been run.
  • Echo. On laptop speakers, only Soniox ears have the echo guard; with other ears, use headphones.
  • Real hardware. Drive mode has not been measured on car hands-free audio. Phone alerts need the app to be running (there is no push server).
  • Code running as you is trusted. It can read Pointe's token and act as you.