Changelog · Verbative

v1.0.0

2026-08-02Initial releaseVS Code

An agentic development extension for VS Code that gives your agents a voice and a shared, persistent memory, driving the Claude Code CLI hands-free. Wake by voice, dictate your prompt, submit, and hear Claude's replies, permission requests, and summaries spoken back to you.

Channels — a team of agents

Run agents in parallel: Group your Claude sessions into channels — one per project, feature, or area — and run several agents inside each, all working at once. Manage it all from the Verbative panel: add a channel or an agent, rename them, and start fresh, resume, or continue any session.
Switch focus in a click: Your dictation and the voice key route to whichever agent you've selected, so switching focus is a single click — and you can mute or spotlight any individual agent's voice.
Shared channel memory: Each channel shares one memory (see below), so a decision made with one agent is recalled by all the others — no re-explaining context. Free includes two channels with up to two parallel agents each; Advanced is unlimited.
External sessions join the channel: Claude Code CLI sessions you start outside Verbative — in any terminal, or in the Claude Code extension's own panels — appear in your Default Channel automatically as External Agents, numbered in the order they arrive. Cards register the moment a terminal session opens (panel sessions appear with their first activity), show the live working state, permission alerts, and the model the session runs, carry the same mute, beep, and pulse controls, and clean up when the session ends. Their spoken output announces the card's name, so you always know which session is talking.
Focus follows you: Click an External Agent's card — or simply focus its terminal — and it becomes your voice-input target, with the same active highlight and the same in-terminal indicator bar managed agents get. Sessions inside the Claude Code extension's panels can't receive dictation, so they never take the voice focus — clicking their card brings the Claude panel into view instead, and their state shows in the channel card and in a status-bar indicator at the bottom of the window, even while the Verbative panel is closed.

Memory — persistent, on-device knowledge

Complete capture, selective recall: The counter-intuitive design at its core: be exhaustive when you capture, selective only when you recall. Guessing what matters at write time is exactly how memory drops the one fact you needed — so Verbative keeps everything and moves the judgment to read time.
Two layers — a lossless ledger under distilled facts: Underneath sits a lossless ledger: a mechanical, no-model record of every command, edit, and test result, so “what did we run” and “when did this test start failing” are exact, replayable facts — not retrieval guesses. On top, an on-device model distills that stream into atomic, typed memories — decisions, conventions, and dead ends (the last as explicit ⚠️ avoid warnings) — deduped and superseded as they change.
Recall fuses both layers: A single recall surfaces only the handful of distilled facts that matter — ranked by semantic search and a reranker — and interleaves them, by time, with the exact ledger record around them: what you decided sitting next to what you actually did (the commands, the edits, the test that flipped), with the causal and temporal links between. One answer that fuses meaning with ground truth, on a strict budget — never a dump of the whole history.
Self-organizing in the background: New facts reconcile against old ones — a contradiction supersedes the stale fact in seconds, a re-confirmation strengthens it. Importance is rated at capture and decays on per-type forgetting curves, frequently-recalled facts grow stronger, memories recalled together become associated, and recurring specifics distill into higher-level insights.
Ask, pin, and manage it: Ask time-anchored questions (“what did we decide last week?”) and manage everything in the Verbative Memory tab — search, sort, pin the always-true rules into every prompt, add or retire entries, and pick the capture mode (Off / On-device / Cloud via the Claude Code CLI). Agents get the same agency through built-in memory tools, including flagging a memory as wrong so it retires on the spot.
Human-readable files you own: Memory lives in human-readable files in your repo — git-ignored by default, with an opt-in toggle to let git track it — in one project pool (shared by every channel, each memory tagged with the channel that learned it) plus a private global scope.
Works with Claude Code CLI and Codex — in the extension or standalone: Memory ships inside the extension and as the standalone verbative-memory packages (pip and npx). Both connect to your agents over MCP: the Claude Code CLI and Codex each get automatic capture and injection, with their own per-agent capture, injection, and context settings. Sign in with your Verbative account email — no keys to manage.
Backfill your history: Memory doesn't have to start empty: backfill reads your past Claude Code CLI sessions — and Codex rollouts — into the same pool, so day one already knows what the last months decided. It runs in windows in the background, with a quiet mode that halves its duty cycle while you work.
Memory Analytics: A dedicated analytics view in the Memory panel: pool health, capture and recall activity over the last 30 days, an activity heatmap, your most-recalled memories, and the breakdown by type — so you can see what your agents actually remember and use.
Fast, local, and proven: On our internal retrieval benchmark the full local stack ranks the right memory first on every query — where a plain keyword baseline almost never does — retrieves in about 100 ms on-device, and makes zero cloud calls.
Part of Advanced: Memory is part of Advanced and starts with a 14-day free trial; if Advanced lapses it simply pauses capturing and recalling — every one of your memory files stays exactly where it is, never touched or deleted.

Hands-free voice control

Spoken command vocabulary: listen, stop listening, hold on, approve, deny, explain, update, open queues, stop claude, repeat, repeat review, skip, pause, resume, and advisory board.
Command your team by voice: Say switch agent or open agents and Verbative reads out a numbered list of your channels and agents, pausing briefly after each option so the numbers are easy to follow. Answer with a number — digits or number words, in English or German — to jump focus, or say cancel. create channel and create agent set up new ones entirely hands-free.
Approve permissions by voice: Hear what Claude wants to do, then approve, deny, or ask it to explain before deciding.
Live Hands-Free panel: A real-time equalizer, an optional working pulse, and an attention beep when Claude needs you. Turning the pulse sound off keeps the working animation visible — it just goes grey, so a silent agent still reads as busy.
Editable voice commands (Advanced): Rename any trigger phrase to whatever feels natural.

Speech & voice, on-device

On-device recognition: Wake-word and command recognition run with whisper.cpp — your voice commands never leave your machine.
Speak any language: Dictate your prompt in any language and Verbative translates it to English entirely on-device with whisper.cpp, then types the clean English prompt into the CLI. Always on, on every plan; the voice commands stay English. On the full Large v3 model it also detects and names the language you spoke.
Two transcription modes: Switchable in a tap: Local (whisper.cpp on-device, full privacy) with a selectable whisper model and a live transcript, or Cloud (the Claude Code CLI's built-in dictation).
Local neural voice output: Spoken summaries, permission prompts, and progress with Kokoro — no cloud TTS. Choose the Claude and mediator voices, adjust playback speed, and set how much of each permission prompt is read aloud.
Bluetooth & audio devices: Bluetooth headset (HFP) support and selectable audio input/output with hot-restart.

Verbative Voice — dictate into any app

A floating button, everywhere on macOS: A standalone macOS app: a floating dictation button that sits above every other window. Click it (or hit the global hotkey), speak, and your words land in the field your cursor is already in — Mail, Slack, your browser, any native app.
Transcribed on your own Mac: Every word is transcribed on-device with whisper large-v3-turbo — your speech never leaves your machine. Dictate in any language and it types clean text right where you left off.
Its own plan, or included with Advanced: Verbative Voice is available on its own subscription, and comes included with Code Advanced at no extra cost.

Coding toolkit

Mechanics: Reusable shell scripts, organized in groups right in the Verbative panel. Save the commands you run all the time (start a dev server with npm run dev, spin up a browser, kick off a build), then run or stop each with a single click. Every Mechanic runs in its own terminal — your real shell and PATH — streams its output live, and stops cleanly. Import an existing .sh file or edit in the built-in shell editor. Free includes up to two Mechanics; Advanced is unlimited.
Advisory Board: A panel of domain advisors reviews every prompt and its result, each speaking in its own voice; the roster is fully yours — add as many advisors as you want. Free includes up to 10 reviews a day; Advanced is unlimited.
Verbative Shot: Press ⌃⌘S (or the panel button) and drag a region anywhere on screen straight to your clipboard. Keep dragging for several in a row (Esc to finish) and they're stacked into one image, so a single paste shows Claude everything at once — a design to build, a render that looks wrong, an architecture sketch to scaffold from. Captures stay on your Mac, nothing is uploaded. Free covers 10 captures a day; Advanced is unlimited.

Platform & plans

Per-workspace isolation: Separate VS Code windows run independent Verbative sessions without cross-talk.
Built for accessibility: Run a complete coding session by voice, keyboard optional.
Runs on your Claude subscription: Works on your existing Claude Pro or Max subscription through the Claude Code CLI — no API key, no SDK. Free includes 10 hands-free loops a day, cloud and local on-device transcription with your choice of whisper model, two channels with two parallel agents each, two Mechanics, the Advisory Board (10/day), and Verbative Shot (10/day); Advanced adds the on-device Memory (with a 14-day free trial), unlimited everything, editable commands, and a 48-hour cached license.