Verbative
Run several Claude Code CLI agents side by side — as native chats, right in your editor: VS Code, Cursor, VSCodium or Devin Desktop (formerly Windsurf).
Every agent works in its own chat: streamed replies, tool cards with their results, and
permission prompts you answer with a click. When you want more, give your agents a voice and
a memory.
- 🤖 Agent Workspace — several Claude Code CLI agents at once, grouped into channels, each in its own chat tile. Start one per feature and watch them all work.
- 🎙️ Hands-free voice — say "listen", speak your prompt, and hear the result spoken back. Approve permissions out loud. Transcribed on-device by whisper, in ~99 languages.
- 🧠 On-device memory — decisions, conventions and dead ends captured from every session and recalled into every prompt, stored as readable files in your repo. Works with the Claude Code CLI and Codex.
It runs the Claude Code CLI you already use, on your existing Claude plan (Pro, Max or
Ultra) — no API key, no extra account.
Works in 20 seconds — no big download
- Install Verbative: from the VS Code Marketplace in VS Code, or from Open VSX in Cursor, VSCodium and Devin Desktop (formerly Windsurf): search "Verbative" in the Extensions view.
- Click the Verbative "V" in the activity bar. A 20 MB runtime sets itself up; channels, the Agent Workspace, Verbative Shot and Mechanics are ready right away.
- Click + New channel, then + Add an agent — it opens as a chat in the Agent Workspace. The free tier activates itself: no sign-up, no API key.
- Want hands-free voice or memory? Click the feature's card: it tells you what gets downloaded and how big it is (the speech engines and models, about 1.3 GB each), then runs on your machine. Nothing downloads until you ask.
Agent Workspace — your agents, side by side
- A native chat per agent — streamed replies, thinking, tool cards with their results, and permission prompts you answer with Allow, Allow always or Deny (or by voice).
- Channels — group agents by project, feature or area; one shared project memory keeps every agent in a channel on the same page without re-explaining context.
- A grid that fits — a lone agent fills the panel; more agents tile automatically until you resize them yourself.
Free includes two channels with up to two parallel agents each; Advanced is unlimited.
The hands-free loop
- A tiny on-device model (whisper.cpp) listens for the wake word — no audio leaves your machine for the trigger.
- Say
"listen" and Verbative captures your prompt.
- Your prompt is transcribed entirely on your machine (the default, free on every plan) — or via the Claude Code CLI's cloud dictation, your choice.
- Say
"stop listening". Claude runs it against your codebase and Verbative speaks a short summary back in a natural neural voice.
More than dictation — the whole loop is voice
- Approve by voice —
"approve" / "deny" answer Claude's permission prompts; "explain" reads back what it wants to do before you decide.
- Stay in the loop — spoken summaries of every change,
"update" for a live status check mid-task, "stop claude" to interrupt, "hold on" to cancel a prompt you don't like, "repeat" to hear the last cue again.
- Command your team by voice —
"switch agent" and "open agents" read out a numbered list of your channels and agents; answer with a number to jump focus. "create channel" and "create agent" set up new ones entirely hands-free.
- Speak any language — dictate in your native language and Verbative translates your prompt to English on-device before it reaches Claude. Always on, on every plan.
- 100% on-device voice — whisper.cpp transcription and the Kokoro neural voice both run locally on your machine, so the entire hands-free loop works without sending audio anywhere.
Memory — your agents finally remember
Your AI coding agent forgets everything between sessions. Verbative Memory fixes that,
fully on-device:
- Complete capture, two layers — a lossless ledger records every command, edit, and test result exactly, while an on-device model distills the stream into atomic, typed memories: decisions, conventions, and dead ends (the last as explicit ⚠️ avoid warnings).
- Selective recall — semantic search plus a reranker surface only the handful of facts that matter, fused with the exact ledger record around them, into each agent's prompt. Precision over volume.
- Self-organizing — contradictions supersede stale facts, useful memories grow stronger, stale ones fade, recurring specifics distill into insights.
- Works with the Claude Code CLI and Codex — inside the extension, and as the standalone
verbative-memory packages (pip / npx) over MCP, each agent with its own capture and injection settings. Backfill reads your past sessions in, so day one already knows what last month decided.
- Yours — human-readable files in your repo (git-ignored by default, opt-in trackable). Manage everything in the Memory tab — search, pin, edit, retire — with a full Memory Analytics view. Nothing ever leaves your machine.
Read more at verbative.de/memory.
- Verbative Shot — press
Ctrl+Alt+S on Windows (⌃⌘S on macOS) from any app and drag a region straight to your clipboard; keep dragging for several in a row (Esc to finish) and they stack into one image, so a single paste shows Claude everything at once. No Accessibility permission needed; captures never leave your machine. Free covers 10 a day.
- Mechanics — save the shell commands you run all the time (
npm run dev, a build, a database reset) as reusable scripts, organized in groups, each run or stopped in its own terminal from the panel. Free includes two.
- Advisory Board — every prompt reviewed in parallel by your panel of domain advisors (Backend, Frontend, Security & Compliance, QA, Product, and DevOps by default), each speaking in its own voice; add, remove, or rewrite advisors freely. Free includes 10 reviews a day.
Fast — because it all runs on the GPU
On-device doesn't mean slow. Everything runs on the GPU — Metal/MLX on Apple Silicon,
Vulkan on Windows — so nothing waits on a network round-trip (measured on an Apple M1 Max):
|
What |
Speed |
| Voice → text |
a 30-second spoken prompt → text |
~0.6 s (≈48× faster than real time) |
| Text → voice |
a spoken reply |
~23× faster than the words take to say |
| Short cues |
"Approved.", "Permission needed." |
~65 ms to first sound |
Transcription uses OpenAI's near-state-of-the-art whisper large-v3-turbo; the voice is
the high-quality Kokoro neural model — accuracy isn't traded for speed.
Private by default
Wake-word and command detection always run on-device. In local mode your prompt is
transcribed entirely on your machine and never leaves it. Your memory never leaves your machine
at all. Our servers only ever see a usage count — never audio, never transcripts, never
memory content. Full policy at
verbative.de/legal/privacy.
Pricing
- Free — 10 hands-free commands a day, forever. Full on-device transcription and voice, two channels with two parallel agents each, two Mechanics, the Advisory Board (10/day), and Verbative Shot (10/day). No sign-up.
- Advanced — €9.99/mo (€4.99 for your first 3 months; or save 30% annually) — everything unlimited, plus the on-device Memory (starts with a 14-day free trial), your pick of whisper model (tiny → large-v3), editable command phrases, a 48-hour offline license cache — and Verbative Voice, the standalone macOS dictation app, included at no extra cost.
See verbative.de/pricing.
Requirements
- macOS 12+ on Apple Silicon (M1 or newer), or Windows 10/11 (x64)
- VS Code 1.80+
- Claude Code CLI installed (
claude in your PATH)
Linux is on the roadmap — track the
changelog. The standalone
Verbative Voice desktop app is macOS-only for now.
FAQ
How is this different from Anthropic's Claude Code extension?
Anthropic's extension puts a Claude Code conversation in a VS Code panel. Verbative is built
for running several agents at once: a grid of native chats grouped into channels with a
shared project memory, a hands-free voice loop (wake word, spoken replies, approvals by
voice), and an on-device memory that carries decisions across sessions. Both run the same
Claude Code CLI, and you can keep both installed.
Does it use my Claude plan?
Yes. Verbative starts the Claude Code CLI you already installed and signed in, so your agents
run on your existing plan — Pro, Max or Ultra, or a Team or Enterprise seat — and count
against its usage limits just as they do in the terminal. Verbative never sees or stores
your Claude credentials and needs no API key.
Does audio leave my machine?
Not by default. The wake word and voice commands are always detected on-device, your prompts
are transcribed on-device by whisper, and replies are spoken by an on-device neural voice.
If you prefer, you can switch dictation to the Claude Code CLI's cloud dictation in the
settings. Our servers only ever see a usage count — never audio or transcripts.
Does it work on Windows?
Yes — Windows 10/11 (x64) and macOS 12+ on Apple Silicon. The Agent Workspace, hands-free
voice and memory work on both. Linux is on the roadmap.
Troubleshooting
Most common first-run issue: a Bluetooth headset stuck without mic access. Use the
built-in mic test in the panel (it shows a live transcript so you can confirm it hears
you), or pick your input device from the dropdown. Full
troubleshooting guide.
Not affiliated with Anthropic
Verbative is an independent third-party product that integrates with the Claude Code CLI.
"Claude" and "Claude Code" are trademarks of Anthropic, PBC.
Made for people who actually code all day.
verbative.de · docs · info@verbative.de
| |