Szept: Say it. Show it. Hand it to your agent.

Author here. I talk to coding agents in Polish with English product words mixed in. ChatGPT and Cursor have good built-in dictation for that, but only inside their own apps; I wanted one light tool that works in Claude Code, mail and everywhere else, and hands screenshots to the agent too. I started with local models (Qwen3-ASR on MLX, then Whisper via WhisperKit for Polish). Recognition was fine; the gigabytes of memory weren't, for an app that stays open all day. Soniox's realtime model worked best for me in Polish, so Szept streams to it on the user's own API key. The parts that might interest HN: - Holding fn over a screenshot pastes the file path plus your spoken sentence into the terminal you came from, so Claude Code sees what you mean. - The cleanup step (an LLM, 3-second deadline, then it pastes the raw text) is told which app the text is going to, including the URL of the browser tab via Accessibility, so Gmail in Brave still counts as mail. - For a while every dictation waited exactly 3 seconds. Soniox ends a stream on an empty text frame, and I was sending an empty binary frame, which it ignores. One-line fix, ~180 ms now. Limits, upfront: no offline mode, audio goes to Soniox, macOS only. It's paid once and ships with the Swift source: https://szept.szyma.co Built mostly with Claude Opus 5.5 in Claude Code, plus DeepSeek Harness on DeepSeek V4.1 Flash for well-scoped tasks. The CLAUDE.md of pitfalls ships with the source. Happy to answer anything.