Daisy
A Mac app that records your meetings and transcribes them on-device, with a local MCP server so Claude can read your calls.

接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
來自發布者
A Mac app that records your meetings and transcribes them on-device, with a local MCP server so Claude can read your calls.

Local-first meeting recorder for Mac. Whisper on-device, your AI for summaries, MCP server for Claude Desktop.
~/Library/Application Support/Daisy (or a folder you pick). Transcription and speaker separation run on-device. No account and no analytics SDK. If you agree, once a day Daisy sends anonymous feature counts — no content, no identifier; the full list is in Docs/usage-statistics.md.*.ingest.de.sentry.io), only when you press Send after a crash — Daisy asks on the next launch and shows the whole report first: where it failed, the app and macOS versions and the Mac model.mydaisy.io, the one-time download of the speech models you chose from huggingface.co (on macOS 26, Apple's speech engine fetches its language files from Apple), and, if you press Download on a local summary model in Settings → Summary → Ollama, that model file from ollama.com.Daisy runs an MCP server inside the app, so Claude Desktop, Claude Code, Cursor or Codex can search and read your meetings.
claude_desktop_config.json runs a few lines of /bin/sh and /usr/bin/curl, both part of macOS (no Node.js, no npx). Claude Code in the Terminal, Codex and Cursor have their own Connect buttons on the same screen.By hand: clients that speak Streamable HTTP use http://127.0.0.1:54321/mcp directly, with the header Authorization: Bearer <access token> (the token is on the same screen, under Privacy → Access token → Copy). For Claude Code:
claude mcp add --scope user --transport http daisy http://127.0.0.1:54321/mcp --header "Authorization: Bearer <access token>"
Claude's own config file takes only a command, not a URL; let Daisy write that entry, or use any stdio-to-HTTP bridge such as mcp-remote.
Tools:
No tool can delete a session, audio or a transcript, or change Daisy's settings.
How the server is fenced:
127.0.0.1 only, so nothing outside your Mac can reach it, and it rejects requests whose Host or Origin is not local.54321; you can change it under Advanced on the same screen.More: https://mydaisy.io/docs/mcp.
Apache License 2.0. You can use, modify and redistribute the code, including commercially, as long as you keep the license and copyright notices. Full text in LICENSE.
Three capture modes, one app:
Remote A / Remote B, with optional mic-side attribution), a summary, action items, and a draft follow-up. Optional extras include periodic screenshots with on-device OCR, a preparation brief built from the agenda and past sessions, evidence-backed progress against the meeting plan, local meeting analytics, and custom meeting apps beyond the built-in list.Around the edges: morning and end-of-day summaries on Home, an opt-in keyboard-layout auto-fixer (retypes text entered in the wrong layout, with undo and per-app exceptions), and provider-returned ChatGPT plan-window usage plus local token accounting for API providers. The UI is localized in English and Russian.
The differentiator: Daisy ships a local MCP server bound to 127.0.0.1 that exposes your sessions as a queryable, actionable data source to any MCP client (Claude Desktop, Cursor, Codex). Because the transcript is already local, Daisy can be a local-only MCP source — something cloud meeting tools structurally can't offer.
Requirements:
Clone and open:
git clone https://github.com/ca33u/daisy-app.git
cd daisy-app
open Daisy.xcodeproj
The Swift Package Manager dependencies (Sparkle, WhisperKit via argmax-oss-swift, FluidAudio) resolve on first project load. Hit Run; the app launches.
Daisy/ → SwiftUI app sources (PBXFileSystemSynchronizedRootGroup)
DaisyTests/ → unit tests
Benchmarks/ → reproducible WER/DER/JER scorer, product runner, and public evidence
Daisy.xcodeproj/ → Xcode project
scripts/
release.sh → end-to-end release: archive → notarize → DMG → sign → Sparkle appcast
release-notes/ → per-version markdown bullets consumed by release.sh
dmgbuild_settings.py → dmgbuild config (Python) for the installer DMG
assets/ → DMG background, app icons
build/ → archive output (gitignored)
RELEASING.md → branch/channel model and the release/promote/hotfix flows
Key services that drive the app:
CoreAudioMicRecorder — CoreAudio mic capture with route-change recovery and the archive .caf writer (replaced the old AVAudioEngine tap to fix route-change/Bluetooth dropouts)SystemAudioCapture / ProcessTapAudioCapture — the remote side of a meeting. The main path on macOS 14.4+ is a Core Audio process tap (the smaller "System Audio Recording Only" permission; it keeps working with Bluetooth output). SCStream loopback is the fallback: on older macOS, before the tap permission is granted or after it is refused, when the tap fails to start, or when it hears nothing that SCStream does. Silent-capture detection and warningsTranscriber / WhisperEngine — WhisperKit on-device transcription with a Silero VAD pre-passParakeetEngine / AppleSpeechEngine — the two alternative dictation engines: FluidAudio Parakeet-TDT (low latency) and Apple SpeechAnalyzer (macOS 26, no model download); Whisper is the defaultDictationPaste — pastes dictated text at the cursor via the Accessibility API, restoring your prior clipboardRecordingSession — orchestrates a session, owns calendar binding and auto-stop schedulingSummarizer — multi-provider LLM dispatch: Apple Intelligence (on-device), ChatGPT account, Anthropic, OpenAI, Google Gemini, Kimi (Moonshot), YandexGPT (Yandex AI Studio), GigaChat (cloud.ru), DeepSeek, Cursor API key, Ollama, LM Studio (local), or an MCP summarizerSensitiveDataProtector — optional on-device pseudonymization/redaction boundary for supported remote summariesMeetingPreparation / MeetingPlanAnalysis / MeetingAnalytics — pre-meeting context, evidence-backed agenda progress, and local call metricsScreenshotCapture — opt-in periodic screenshots of the meeting window with Vision OCR; screen text flows into the transcript and summaryPreMeetingBrief / MorningBrief / EndOfDaySummaries — local briefs assembled from your calendar and past sessions, plus an evening digestVoiceProfile — opt-in personalization learned from your dictations (and, optionally, your mic side of meetings)LayoutAutoFix — opt-in keyboard-layout auto-correction via a CGEvent tap, with undo and per-app exceptionsTokenLedger — local token-usage accounting per cloud provider, shown on HomeMCPServer — the local MCP server on 127.0.0.1; exposes ten tools (five read, two local edits, three opt-in actions) to Claude Desktop / Cursor / CodexVoiceMemoScanner / VoiceMemoIngestor — opt-in, on-device import of Apple Voice Memos to Markdown transcriptshttps://mydaisy.io/appcast.xmlBenchmarks/ contains the product-pipeline runner, a neutral
standard-library scorer for WER/CER/DER/JER, fixtures, and published raw
evidence. The first public baseline is AMI ES2004a: Daisy 1.0.7.59 detected
4/4 speakers with 15.68% DER and 20.28% JER at a median 0.122× real time
across three warm runs on an M4 MacBook Air. It is one reproducible
diarization case, not a general accuracy
claim; Humla and OpenWhispr remain unscored until their raw output exists for
the exact same audio. See the methodology and publication gate
and the public evidence.
DAISY_AUTO_PUSH=1 ./scripts/release.sh <shortVersion> <buildNumber> [stable|beta]
Beta is the default channel from main; stable is promoted from a soaked beta with ./scripts/release.sh promote <version> (no rebuild). Six steps: archive → export → notarize → DMG → publish to the daisy-web repo → inject an <item> into appcast.xml and commit. Vercel auto-deploys the site within a couple of minutes. Every stable version also becomes a GitHub Release with the same DMG (./scripts/release.sh github-release <version>). Full branch/channel model and the hotfix flow are in RELEASING.md.
Release notes for each version go in scripts/release-notes/<shortVersion>.md as a flat markdown bullet list (- one line per change). The script extracts those bullets and embeds them in the appcast <description> so Sparkle shows them in its update sheet.
SECURITY.mdDownloaded when first needed — none of them ships in the installer. Their authors keep their rights; the credit doesn't imply endorsement. Full list with sources: mydaisy.io/licenses. The licenses and notices of everything Daisy for Mac links and downloads ship inside the app as Acknowledgements.txt (generated by scripts/acknowledgements.py).
wespeaker_v2): fixed input shapes, a fused filter-bank front end, 8-bit palettized weights. Both speaker models: FluidInference/speaker-diarization-coreml (see its NOTICE.md).Local-first meeting recorder for Mac. Whisper on-device, your AI for summaries, MCP server for Claude Desktop.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Daisy App 透過 stdio 在本機執行。
上次檢查時間:2026-11-17。 Local-first meeting recorder for Mac. Whisper on-device, your AI for summaries, MCP server for Claude Desktop.