Daisy

On-device transcription is powered by WhisperKit, running on the Mac's Neural Engine, which delivers fast and accurate speech-to-text comparable to cloud services. To eliminate hallucinations caused by silence, Daisy employs a Silero VAD (Voice Activity Detection) pre-pass that filters out non-speech segments before transcription begins. Speaker diarization is handled by Pyannote, which labels remote voices as "Remote A", "Remote B", and so forth, with an optional mic-side diarization feature that attributes your own display name when speaking. A key differentiator is Daisy's two-channel approach: your microphone and the call's output are captured on separate channels, so diarization only needs to sort voices within the remote side—reducing errors significantly. Furthermore, when you hit stop, Daisy runs an offline re-diarization pass using FluidAudio (from the pyannote family), achieving approximately 10–15% diarization error on standard benchmarks, compared to the 26–50% error typical of real-time streaming solutions. If the system misidentifies a speaker, a one-click rename allows you to correct it, and Daisy remembers that voice going forward.

All transcripts are saved as plain Markdown files with YAML frontmatter in a user-selected folder—commonly an Obsidian vault or iCloud Drive—giving users full ownership and easy access through any text editor. Daisy never holds your data on a server; each recording becomes a simple .md file you can inspect, copy, delete, or version-control. The most transformative feature is the built-in MCP (Model Context Protocol) server, which listens on 127.0.0.1 and exposes nine tools to AI clients like Claude Desktop and Cursor. With one click, Daisy writes the required configuration into Claude Desktop's claude_desktop_config.json, preserving any existing MCP servers. The tools include five read operations—list_sessions, get_session, search_sessions, list_folders, list_destinations—and four action operations: re-summarize, retitle, rename a diarized speaker, and route a session to a destination. This allows your local AI to search, retrieve, and act on meeting transcripts without anything leaving your Mac. The server stops automatically when toggled off.

After recording, Daisy processes audio entirely on-device: Silero VAD filters non-speech, WhisperKit transcribes on the Neural Engine, and Pyannote performs offline diarization to assign speaker labels. The result is a Markdown file stored in your chosen folder. If AI summaries are enabled, Apple Intelligence (offline) or a bring-your-own-key provider (Anthropic, OpenAI) generates a summary appended to the transcript. From there, you can manually or automatically push sessions to destinations like Notion, Linear, Attio, Slack, or a webhook. The MCP server exposes nine queryable tools to Claude Desktop and Cursor, turning your meeting history into a live data source without any file uploads. Daisy requires no account, no subscription, and no internet for core recording and transcription—just the app and permission to record system audio.

A typical use case is a consultant recording client interviews. Daisy captures the conversation locally, transcribes it, and automatically routes the transcript—along with AI-generated action items—to a Linear project or a Notion database. The consultant can later ask Claude Desktop, "What were the key deliverables from the Q3 review meeting?" and the AI lists_sessions on the local MCP server, retrieves the relevant transcript, and summarizes the outputs—all without any data leaving the Mac. For remote teams, daily standups can be recorded and shared via Slack webhooks, with corrected speaker labels for clarity. Privacy-sensitive organizations, such as law firms or healthcare providers, can deploy Daisy without a DPA because there is no cloud storage or processing. Quick note-taking is handled by voice notes mode, and dictation mode allows users to speak text directly into any application. The outcome is a streamlined workflow where meeting content is captured, organized, and actionable—while staying entirely under the user's control.

Daisy is built for privacy-conscious professionals, including consultants, developers, compliance officers, and anyone who wants to own their meeting data. It runs exclusively on Apple Silicon Macs (M1 through M4) with macOS 14 or newer, leveraging the Neural Engine for fast on-device transcription. The app is open source under Apache 2.0 and fully auditable. During the beta, Daisy is free to use; after launch, it will be available as a one-time lifetime purchase—no monthly subscriptions, no per-meeting fees, and no per-user bills. This model contrasts sharply with cloud-based meeting tools that charge $8–25 per month forever. Users can integrate with any AI provider, from local Ollama and LM Studio models to Anthropic Claude and OpenAI, using their own API keys. The bottom line: Daisy offers a complete meeting recording, transcription, and summarization solution that remains on your Mac, giving you full data sovereignty, transparent pricing, and immediate compatibility with your existing tools.