Get it for macOS

Hear one spoken sentence when an agent needs you, and see every waiting thread in one list. When a thread asks a question, waits for a plan or a permission to be approved, finishes its turn or fails, the bb app says which thread it is and what it needs, on whatever page you are on. A Herald page in the sidebar lists every thread waiting on you with its reason and sentence, and the same sentence sits above that thread's composer with a Read again button until you answer. Settings choose which kinds of event are announced, the voice and speed, and where to speak: the desktop app, a browser tab, or the mobile app while it is open. By default the sentence is built from the event itself, with no model or external service; optionally, Claude Code, OpenAI Codex, Gemini CLI or a command of your own, installed and logged in on the Mac running bb, writes it instead, for one short model turn per announcement. The default voice needs bb running on a Mac, and elsewhere each device's own browser voice speaks. Requires bb 0.47 or later.


Overview

View source

Each sentence is written for the ear: which thread it is, what kind of event, and the question, the command or the start of the reply, kept short. It is spoken on whatever page of bb you are on, so you know which thread to go back to without watching the sidebar.

What you get

  • When an agent asks a question, waits for a plan or a permission to be approved, finishes its turn or fails, the bb app says one sentence about it — what is being asked and the choices, the command, or the start of what was done.
  • A Herald page in the sidebar listing every waiting thread with its reason and sentence, and a count on its sidebar entry. Tap a row to open the thread, or Read again to hear it again.
  • Below them on the same page, the latest sentences from every thread in every project, 20 by default and up to 50, each opening its thread.
  • The sentence above the waiting thread's composer, with a Read again button, until you answer.
  • Settings for where to speak (desktop app, browser tab, mobile app), the voice and speed, and which kinds of event are announced.
  • A choice in each thread's menu: Announce, Questions only, or Silent. A worker without one follows the thread that started it, so silencing a coordinator silences its workers. On bb 0.48, set it for several threads selected in the sidebar at once.
  • Optionally, a model writes each sentence: Claude Code, OpenAI Codex, Gemini CLI or a command of your own, installed on the Mac running bb, runs once per announcement with no tools or read-only. Off by default; the plain sentence is spoken whenever the tool fails or is slow.

Beside bb's push notifications

bb's own push notifications tell you a thread finished, failed or waits for input, with its last words. Herald adds a sentence to be heard, speech on every client, and one list of what is waiting. The two can run together.

What it needs

  • bb 0.47 or newer.
  • No account or external service for the plain sentence: it is built from the event itself. A model-written sentence needs the chosen command-line tool installed and logged in on the Mac running bb, and each announcement costs one short model turn there.
  • For the default voice, bb running on a Mac, whose say voices render each sentence. Otherwise each device's own browser voice speaks.
  • A browser tab speaks after you press Test voice once. The mobile app speaks only while it is open on screen, after one press of Test voice, and only once you switch it on.

More from Gustavo Ambrozio

ChatGPT Realtime ConnectionRun Voice Mode's voice chats on your ChatGPT subscription instead of a separately billed OpenAI API key. It hands Voice Mode the ChatGPT sign-in that the Codex CLI already keeps on the machine running the bb server, reads it again for every chat, and never renews or rewrites it. Only the plugins you allow, Voice Mode to begin with, receive the token; the bb app, the command line and agents in threads never do. This is an undocumented use of the subscription that OpenAI could stop accepting at any time, nobody has checked whether OpenAI's terms allow it, and voice chats count against your ChatGPT plan. Requires bb 0.46 or later, Voice Mode 0.1.9 or later, and the Codex CLI signed in with Sign in with ChatGPT on a plan that includes Codex.26Diff ReviewComment on a thread's changes line by line, then send the comments to its agent, which fixes the code and answers each one. A Review tab beside the thread shows the diff of its uncommitted, staged or unstaged changes, its branch, one commit or a range, and each comment follows its code as the code moves. The same tab reviews a GitHub pull request: answer its threads, mark files viewed in step with GitHub, and submit your comments as one review, posted as the GitHub CLI's signed-in account. An AI review leaves comments of its own, @reviewer answers a question inside a comment thread, and Fix in a new thread hands the comments to another agent; these runs start as threads under the one being reviewed, in the permission mode set in the settings (bb's Approve for me by default, or Accept edits), and in full permission mode with a provider that lacks that mode, such as opencode. Follows the review loop of the open-source diffity, credited in the plugin's third-party notices. Needs bb 0.48 or later with bb's Git plugin on, and the signed-in GitHub CLI (gh) for pull request review only; it is built on experimental bb plugin APIs that may change between bb releases.NewDreamKeep every Claude Code project's memory consolidated overnight, as Claude Code's own auto-dream would if it ran in bb. Each dream is a hidden thread that can read anything but write only that project's memory files, and a project's first dream is a dry run you review as a diff, then apply or discard. The memory is backed up before every change, and the Dream page shows each project's memory, why it is or is not due, the dry runs waiting for you and the history of dreams. The schedule is off until you turn it on. Each dream is one model run on your Claude Code account; requires bb 0.49 or later, macOS or Linux, and Claude Code signed in on the machine dreams run on.NewFirstMate CrewHand one agent your to-do list and it runs a crew of worker threads, each in its own git worktree, until the work lands. The first mate is a pinned thread in bb's own chat that writes each worker's instructions, supervises it, and brings back branches, pull requests, findings and only the decisions that are yours. A FirstMate tab shows the crew as a board of Queued, Working, Blocked, Parked, Done, Failed and Idle, with Steer, Interrupt, Relaunch and End on every card. A built-in watch tells the first mate within about five minutes when a pull request is merged, reviewed or changes check status. Requires bb 0.47 or later, git, and the logged-in gh CLI for the pull request watch; a capable model such as Claude Sonnet is recommended for the first mate.14

More in Notifications 10

Know when work finished, failed, or needs attention.

View all