Szept

Product Introduction

1. Definition: Szept is a native macOS productivity application (macOS 14+) designed for multimodal interaction with AI agents. It functions as a lightweight system-wide dictation tool, AI-powered text processor, screen capture utility, and context bridge for AI coding assistants.
2. Core Value Proposition: Szept exists to streamline the workflow of developers, technical writers, and knowledge workers by enabling them to dictate code and instructions in over 60 languages, instantly capture and annotate screen context, and seamlessly hand off both voice and visual data to AI agents like Claude Code and Cursor without leaving their primary development environment.

Main Features

1. Cloud-Powered Multilingual Dictation: Szept uses the fn key as a push-to-talk microphone. Audio is streamed in real-time via TLS 1.2+ to the Soniox stt-rt-v5 speech-to-text API, chosen for its low-latency multilingual support. The app itself contains no local speech model, resulting in a ~180ms latency from key-up to text insertion and a tiny 4.9 MB install size. It supports dynamic language switching mid-sentence.
2. Context-Aware AI Text Cleanup: Activated with fn + ⌘, this feature sends dictated text to a configurable AI endpoint (DeepSeek by default) for post-processing. It performs semantic cleanup—removing filler words, fixing punctuation—and formats output based on the target application's context (e.g., adding email signatures for Mail, preserving single-line format for Slack, keeping exact commands for Terminal).
3. Agent-Centric Screen Capture & Handoff: The ⌘ ⇧ A shortcut triggers a screen freeze for region/window selection. Beyond standard annotation, it includes productivity tools like text extraction (OCR), QR code reading, and "Scrollshot" for full-page captures. The ⌥ ⇧ R shortcut records screen regions with system audio and microphone. The core innovation is the "handoff": with a screenshot or recording active, holding fn and speaking appends both the transcribed audio and the local file path directly into the active window of an AI agent like Claude Code.
4. Local, Searchable History & Media Library: All dictations, screenshots, and recordings are stored locally in a searchable timeline (⌘ ⇧ H). The search function is language-agnostic, finding "wdrozenia" from the transcript "wdrożenia." The Media tab provides a contact-sheet view. This creates a private, on-device knowledge base of all interactions.

Problems Solved

1. Pain Point: Fragmented workflows for AI-assisted development. Developers must manually switch between dictation apps, screenshot tools, and their AI agent, then copy-paste or drag-and-drop files, breaking concentration and slowing down iterative problem-solving.
2. Target Audience: AI-Powered Developers using Claude Code, Cursor, or Codex; Multilingual Technical Professionals who think or code in non-English languages; Product Managers & Technical Writers who frequently document bugs or create instructions with screen captures.
3. Use Cases: Debugging UI Issues: Recording a screen region with a bug, dictating "The submit button overlaps the footer on narrow viewports," and handing it directly to Claude Code for a fix. Multilingual Stand-ups: Dictating daily notes in Polish, switching to English for code snippets, and having the text cleaned and formatted for Slack. Creating Tutorials: Using scrollshot to capture an entire documentation page, annotating steps, and recording a voiceover, all within one tool before handing assets to an AI for draft assembly.

Unique Advantages

1. Differentiation vs. Built-in macOS Dictation: Szept offers superior multilingual accuracy and mid-sentence switching, AI-powered contextual cleanup, integrated visual capture, and direct handoff to coding agents—features absent in the native OS tool. Vs. Local Whisper Apps: Szept's cloud-based Soniox engine avoids the disk/memory footprint of local models (2.9GB for Whisper large-v3 vs. 4.9MB for Szept) and offers faster, more accurate multilingual transcription without taxing the CPU.
2. Key Innovation – The "Agent Handoff" Protocol: Szept's defining innovation is its ability to bundle a spoken query with a locally referenced screenshot or video file path and inject it directly into an AI agent's input field. This creates a seamless, keyboard-free pipeline from observing a problem on screen to requesting a solution from an AI, dramatically reducing context-switching overhead.

Frequently Asked Questions (FAQ)

1. How does Szept's speech recognition work and is it private? Szept streams audio to your own Soniox API account using TLS encryption. Soniox states it does not store audio/transcripts from real-time streams for training and holds SOC 2 Type 2 / ISO 27001 certifications. Szept itself runs no servers, ensuring your voice data is processed only by the provider you control.
2. Can Szept work completely offline? No. Core dictation and AI cleanup require an internet connection to utilize the Soniox and DeepSeek (or alternative) APIs, respectively. However, screen capture, recording, media management, and history search function fully offline.
3. What is included in the "source code" purchase? You receive the complete Swift project in a private GitHub repo, including the app, 638 passing tests, build scripts, and a proprietary "Agent Kit." This kit contains 78 documented pitfalls (e.g., macOS sandboxing quirks), 12 implementation plans, and 5 design specs to enable AI agents to effectively understand and modify the codebase.
4. Is Szept a subscription service? No. It is a one-time purchase with a perpetual license. The founding price of $9.99 includes 12 months of updates; after that, you can choose to renew updates annually for $9.99 or continue using the last version you own indefinitely.
5. How does the AI text cleanup handle data privacy? If enabled, cleanup text is sent to the configured endpoint (default: DeepSeek, which may store data in China). For maximum privacy, users can disable cleanup, use dictation only, or point the feature at a self-hosted or enterprise OpenAI-compatible API endpoint they trust.