Readback

Hear your AI coding agent's replies in a Speechify voice, with the words highlighted as they are spoken.

Claude Code works through a task. Readback reads the notes Claude writes between tool calls as they land, then condenses the finished reply to a few spoken sentences, all in the VS Code sidebar with each word lit as it is said. The full reply is one click away. Click a word to jump to it. Step by sentence. Cycle the speed. Pick any voice your Speechify key can render, including your own clone, and hear a sample of each before you choose.

Install

Paste this into Claude Code:

Install the Readback VS Code extension, then run its "Install the Claude Code hook" command and tell me when it is done.

Or by hand: install the extension, then either answer the prompt that appears on first start, click Readback in the status bar, or run Readback: Set Speechify API key from the Command Palette. Then press Install the hook in the Readback view.

You need a Speechify API key. Create one at https://platform.speechify.ai/api-keys?utm_source=github&utm_medium=readme&utm_campaign=readback. Readback keeps it in VS Code's secret storage and every reply you listen to bills your own workspace.

After a reload, click once anywhere in the VS Code window before you expect to hear anything. The browser inside VS Code keeps audio silent until then, and the player shows a card until it has played once. Replies that arrive before that queue up.

Claude Code

Readback is also a Claude Code plugin: a band above the prompt shows the reply and lights each word as it is spoken. It needs Claude Code 2.1.275 or later, in the terminal or the desktop app. In T3 Code and other hosts that draw no band, /readback play and the other words below take the keys' place.

/plugin marketplace add Speechify-AI/readback
/plugin install readback@readback

From a shell, claude plugin install readback --marketplace Speechify-AI/readback adds the marketplace and installs in one step.

The install dialog asks for your Speechify API key. Create one for free at https://platform.speechify.ai/api-keys?utm_source=github&utm_medium=readme&utm_campaign=readback. The key is stored as a sensitive plugin setting and billed to your own workspace; change it any time in /plugin.

• /readback on and /readback off turn Readback on or off; bare /readback prints the status (on or off, voice, speed, what is playing). Turning on checks the key against the live API. Nothing plays or bills until you press play. On or off is read when a session starts, so turning it off in one terminal leaves sessions already open on until they restart.
• /readback play, pause, resume, stop and full do what the band's buttons do, for a host that draws no band (T3 Code and other SDK hosts); each word does what it says and nothing else. There a new session sleeps until /readback is typed once, since nothing tells it from a script's headless run, so its first reply is not heard. With auto on it starts at once instead, and so does any headless claude -p run on the machine while autoplay stays on.
• /readback again plays the last reply from its start. The render is cached, so a replay costs nothing.
• /readback auto on plays each reply as it lands, as if you had pressed play; off by default, and every reply then renders and bills.
• Typing the next prompt stops whatever is playing, as in the extension.
• With autoplay on, the notes Claude writes between tool calls are read as they land, as the extension reads them; the finished reply follows. Stop (2) during a turn keeps its later notes quiet; the reply still plays.
• One alert is spoken, on by default: a permission prompt that has waited a few seconds for you ("Claude is waiting for your permission to run npm install"). /readback alerts off silences it. It is one short render per tool call, cached after the first time.
• /readback voice opens a voice picker in the band: a search field and the matching voices, featured first. The band lists 64 at a time; type to narrow. /readback voice harper (a name, language or tag) picks the first match by word, where no band draws. The choice is this project's; /readback voice default harper sets the voice for projects without one, and /readback voice reset drops this project's. /readback voice random draws a voice in your language that no project uses yet, featured first; run it again for another. A voice per project tells sessions apart by ear.
• Each reply opens with the project's folder name, as in the extension, so you know which session is speaking.
• /readback speed 1.25 sets the speed (0.75 to 2), and the band's speed button steps through the same values. Speechify renders the rate into the audio, so a changed speed re-renders: what was loaded is dropped and 1 plays the turn again at the new rate.
• Keys 1, 2 and 3 play or pause, stop, and play the full reply. They work only while the prompt is empty: a "1" typed as the first character of a message goes to Readback, not to Claude.
• While Readback is on, the plugin adds a short section to the system prompt asking Claude to end each long reply with a spoken note, an HTML comment holding one to three sentences for the ear. That note is what is read, at once, with no second run. A long reply without one is condensed with claude -p on your own Claude subscription when the turn ends. The full reply is one press of 3 away either way. Condensing needs the claude CLI and has been checked on macOS only; on Windows a reply without a note is read in full.
• Several Claude Code sessions share one speaker. Pressing play in one pauses another, which resumes once the first is done. A reply that lands while another session is speaking waits its turn instead of cutting in; the band says "waiting for another session to finish" meanwhile.
• Renders are cached under ~/.readback/claude-code/renders, so replaying costs nothing.

The VS Code extension and the plugin are independent; use either or both. With both, a reply is read once: the plugin stays quiet when a VS Code window with that project open is listening, and reads as usual when VS Code is closed or open on another project.

How it works

• Five hooks in ~/.claude/settings.json, MessageDisplay, PreToolUse, Notification, UserPromptSubmit and Stop, post each assistant message, each tool call, each permission prompt left waiting, each prompt of yours and each finished turn to Readback. The hook is ten lines of POSIX shell and curl (the same lines in PowerShell on Windows). It needs no runtime on your PATH and never decides anything itself.
• Codex CLI can be a second source: a Stop hook in ~/.codex/hooks.json posts each finished turn the same way. Codex has no per-message hook, so its notes between tool calls are not read. See Codex below.
• A message followed by a tool call was a note written before it, and Readback reads it as it stands while the tool runs. A message followed by Stop is the finished reply and gets condensed, unless it is under 500 characters already, which is read as it stands. Turn the notes off with readback.progress.
• A permission prompt that has waited about six seconds (Claude Code's own threshold for the permission_prompt notification) is read at once, ahead of whatever is playing: "Claude is waiting for your permission to run npm install, in readback." The tool call comes from the PreToolUse that preceded it. Whatever was playing carries on afterwards. Turn it off with readback.alerts.
• When you submit your next prompt, Readback stops reading: the reply it was on has been read. A reply still being condensed when you type lands in the list without playing.
• Each turn opens with a one-word lead-in, the project's name, that renders once and plays from the cache, so you hear that a turn has started within a few milliseconds while its first paragraph renders. The time is shown next to it, not spoken.
• Readback condenses the reply with claude -p on Sonnet, on your own Claude subscription, so it holds no model key. The run loads no settings and is marked so the hook ignores it, which is what stops a reply about a reply. If claude is missing or the run fails, you hear the full reply instead. Turn it off with readback.summarize.
• Readback flattens the text to plain prose (code blocks become "code omitted", a table of up to four rows its header and value pairs, a longer one a note, links their text), rewrites the few things a voice gets wrong (a bare URL is "a link", app.ts:42 is "app.ts, line 42", a commit hash its first four characters, v0.9.0 "version 0.9.0", >= and the like as words), splits it into paragraphs, and synthesizes each one through /v1/audio/speech as the player reaches it. What is shown is the same text.
• A reply that ends with its own spoken note (an HTML comment the Readback plugin asks Claude for while it is on, <!-- readback: ... -->) is read from that note; no condensing run is made for it. The comment is never shown.
• Speech marks come back with character offsets into the exact text sent, and that is what drives the highlight. No guessing, no token alignment.
• Renders are cached on disk by text, voice and model. Replaying is free. Changing voice re-renders.

Commands

Command	What it does
Readback: Set Speechify API key	Checks the key against the live API before storing it
Readback: Choose voice	Opens the picker in the panel: featured voices first, then every voice grouped by language, with search, filters by gender, clone or tag, and a sample of each. Saved for the open project; set readback.voice in user settings for the default
Readback: Install the Claude Code hook	Adds Stop, MessageDisplay, PreToolUse, Notification and UserPromptSubmit hooks, leaving the rest of the file alone
Readback: Remove the Claude Code hook	Removes exactly that hook
Readback: Install the Codex hook	Adds a Stop hook to ~/.codex/hooks.json. Also a row in the settings panel when a ~/.codex folder exists
Readback: Remove the Codex hook	Removes exactly that hook
Readback: Catch me up	One briefing of every listed reply you have not heard. The history button in the bar does the same
Readback: Read selection	Reads the editor selection. Also in the editor context menu
Readback: Stop	Stops playback and clears the queue
Readback: Clear the list	Forgets every listed reply. The trash button in the bar does the same

Settings

Setting	Default	Meaning
readback.autoplay	true	Play each new reply as it arrives; off lists them and waits for play
readback.summarize	claude	Condense replies with claude -p, or off to read them in full
readback.progress	true	Read the notes Claude writes between tool calls as they arrive
readback.alerts	true	Say when Claude has been waiting for your permission for a few seconds, ahead of whatever is playing
readback.voice	harper_32	Speechify voice id. Per project when set from the picker. Unset, and Claude Code's language is one the model speaks, a voice in that language is used
readback.model	simba-3.5	Speechify model. On an English-only model (simba-3.2) other voices render on simba-3.0 instead
readback.speed	1	Playback rate. Per project when cycled from the player
readback.minChars	80	Replies shorter than this are skipped
readback.maxChars	4000	Longer replies are cut at a sentence and end with "and more"

Languages

Readback speaks whatever language Claude writes in, as long as the voice speaks it too. On simba-3.5, the default, every voice speaks its 48 languages. On simba-3.2, which is English only, a voice in another language renders on simba-3.0, which speaks German, Spanish (Spain and Mexico), French, Italian and Brazilian Portuguese. The voice's locale goes with every request.

To have Claude answer in Spanish, add language to ~/.claude/settings.json, or to a project's .claude/settings.json for that project only:

{
  "language": "spanish"
}

Claude Code applies it straight away. Readback reads the same setting:

• If you have not picked a voice, it uses one in that language.
• If you have, it keeps yours. The voice picker lists that language first and says when your voice speaks a different one. Pick a voice from that group.

The condensed summary is written in the reply's language, so it matches what Claude wrote. The catch-up lead-in ("Catching up on 3 replies") is still English. If Claude writes in a language no voice speaks, such as Japanese, the reply is read by your voice anyway and will not sound right.

Several windows

Each window shows and speaks only its own project's turns. A turn belongs to a window when the folder Claude Code ran in is inside one of its workspace folders, or contains one (running Claude at a monorepo's root while the window has a sub-project open is the same project). A turn that matches no open window is not read.

Catching up

A reply counts as heard when it started playing while this VS Code window was focused. One that played to an empty room, or that waited because autoplay was off, stays unheard. Catch me up (the history button, or the command) takes every unheard reply in the list, oldest first, and has claude -p brief them together: what landed, what is still open, what waits on you, grouped by project. The briefing is listed as its own turn and plays; the replies it covers stay in the list and count as heard. Without claude, or with readback.summarize off, the same replies are read one after another instead. At most 20 are read in detail; older ones are counted in the lead-in.

Codex

Open the settings panel and press Install on the Codex CLI row, or run Readback: Install the Codex hook. This writes one Stop entry to ~/.codex/hooks.json, next to whatever is already there. Codex runs a new hook only after you have trusted it once: open Codex, run /hooks, and approve Readback. Until then nothing arrives. Condensing still runs through claude -p when it is installed; without it Codex replies are read in full.

The player

Play, pause, back and forward a sentence, stop. Space, left and right do the same when the player has focus. The progress line tracks the current paragraph. The autoplay toggle decides whether a new reply plays at once or waits, marked with a dot, for you to press play. The speed pill cycles from 0.75× to 2×; the voice pill opens the voice picker. Each turn has its own play button, and a condensed turn has a "Full reply" section with one of its own.

Limits

• Claude Code and Codex CLI are the sources. The Claude Code hook fires for sessions in a terminal and in the Claude Code panel in VS Code.
• On Windows the hook is a PowerShell script run through powershell.exe with no shell in between. It has not yet been tried on a Windows machine by us; if you run it there, an issue with what you saw would help.
• Read-along needs speech marks, which the streaming endpoint does not return, so replies are read paragraph by paragraph at up to 2,000 characters per request.

Develop

npm install
npm run typecheck && npm test
npm run build          # dist/extension.js
npm run package        # readback-<version>.vsix

Press F5 in VS Code to run the extension in a development host.

The Claude Code plugin is .claude-plugin/, hooks/ and mod/. Try it with claude --plugin-dir . and check it with claude plugin validate ., which npm test runs when claude is on your PATH. It ships from a tree of its own, without the extension's package.json and lockfile: node scripts/plugin-tree.mjs --out <dir> writes it, and --commit commits it to the plugin branch and tags it readback--v<version>.

MIT. Built by Speechify. The player's icons are VS Code's codicons (CC BY 4.0).