You can talk three times faster than you type — and prompts are long. Prodius is a system-wide voice keyboard: dictate prompts, commit messages and reviews into Cursor, Claude Code or any app, clean and punctuated, with your identifiers spelled right. And when an agent is truly blocked — a permission, a decision it can't make alone — it can reach you as a popup you answer with a click or your voice.
Working with agents means describing intent in full sentences, all day. Hit the shortcut, talk, hit it again — Prodius types clean, punctuated text wherever your cursor is.
Dictate prompts into Cursor chat, the Claude Code terminal or Windsurf — same hotkey everywhere, no plugin per tool.
Handles mixed-language speech, library names and identifiers. A personal dictionary keeps your repo and project names spelled right.
Commit messages, PR descriptions, code review comments, Slack and Jira replies — anywhere you can type, you can talk.
Running agents in three terminals? Lock dictation to one — a green border marks the target and stays on it, so your answer never types into the wrong session.
Most agent chatter should stay in the terminal. Prodius runs an MCP server your agents call only when they're actually blocked — a permission, a destructive action, a decision they can't make alone. You define when that is, in plain English.
A permission prompt, a destructive action, a decision it can't make alone. Instead of stalling silently — or guessing — it calls ask_user.
A popup appears over whatever you're doing — browser, IDE, another project. With the agent's name, the context, and the options.
Pick an option, or press the shortcut and answer by voice. The agent gets your answer and keeps going. You never opened the terminal.
Plain English in your project config — no settings UI, no YAML. Different repo, different rules.
…and any tool that speaks MCP: build scripts, CI pipelines, custom agents.
Free-form answers, not just buttons. Explain the constraint, dictate the commit message, redirect the approach — hands-free.
Dangerous command? Scope change? The agent asks before acting, and your answer is logged in the session — not lost in a scrollback.
Run Claude Code on one repo and Cursor on another. Each popup is labeled with its agent and task, so parallel sessions stay sane.
Your reply goes straight back into the agent's context as if you typed it — logged in the transcript, not lost in a scrollback.
You're dictating into repos, terminals and internal tools. Here is the honest data path — the same one we'd want as users.
Audio and transcripts pass through our inference gateway in memory and are never written to our servers. No dictation archive exists on our side.
Transcription and cleanup run through provider APIs under terms that exclude model training. Your code names and prompts don't teach anyone's model.
Context capture auto-blocks password managers and private browsing windows, and never touches your clipboard to read a selection.
On the Practice plan, all processing is pinned to EU infrastructure with a DPA. Free and Pro use vetted US providers under no-training API terms.
Full details: privacy policy and subprocessor list.
Cursor, Windsurf, Codex and raw mcp.json configs, the agent usage protocol, and troubleshooting: full setup guide →
Every new account starts with 14 days of Pro — unlimited dictation, no card. After that: up to 2,000 words per week, free, no end date — agent popups and approvals are unlimited on every plan. Keep unlimited dictation and AI actions for €15/month (€12 billed annually).
Works on Windows and macOS, with every MCP-capable agent. Agent popups and approvals are free and unlimited; the free tier includes 2,000 dictated words per week.
Download free