Local-first macOS meeting-notes app that records or imports audio, performs on-device transcription and summarization, identifies speakers, generates notes/action items, and exposes a read-only MCP server for Claude.
mojiroku is a macOS meeting-notes app designed to keep the core pipeline on the user's Mac. It can record microphone and system audio, import audio files, transcribe locally, separate speakers, generate summaries/meeting notes/action items, search the local meeting archive and export to Notion, Slack or local formats. A built-in MCP server lets Claude Desktop or Claude Code search and read meeting notes, with the MCP surface described as read-only. 1
The implementation is unusually inspectable for this category. The public repository documents a Tauri/Rust desktop architecture using whisper.cpp for speech-to-text, sherpa-onnx for diarization and llama.cpp/local models for summarization, with optional BYOK OpenAI or Anthropic summaries. The repository is published under AGPL-3.0-or-later. 2
The app and local AI path are free, require no account and have no stated recording-minute or storage subscription cap. Optional cloud summaries use the user's own provider API key and therefore incur provider charges directly. 1
The accessibility trade-off is hardware: mojiroku currently requires an Apple Silicon Mac running macOS 11 or newer; Intel Macs and Windows are not supported. Initial model downloads total several gigabytes even though the application installer itself is small. 12
The strongest evidence is first-party rather than independent. The current site gives unusually specific privacy boundaries: recordings, transcripts, summaries and voiceprints are stored locally by default; local features work offline after models are downloaded; transcript text leaves the device only when the user opts into cloud summaries, while chosen integrations can transmit exported content. API keys and integration tokens are described as stored in macOS Keychain. 1
The public source repository materially improves inspectability of those architectural claims, but source availability is not the same as independent security or quality validation. The repository itself says the next phase includes broader beta use and real-meeting quality evaluation. 2
The fixed 50-source AiToolMap panel found no eligible exact-product numeric user-rating evidence: 44 sources produced no exact-product result and six were inaccessible in the audit path. Missing evidence is neutral. No AiToolMap hands-on test was performed, so transcription accuracy, diarization accuracy, summary quality, latency and battery impact are not claimed here.
mojiroku is strongest for individual Mac users and small teams that treat meeting audio as sensitive and prefer local processing over bot-based cloud transcription. It also fits users who want their meeting archive queryable from Claude through a constrained read-only MCP interface. 1
It is a weaker fit for Windows/Intel-Mac environments, centrally administered cross-platform enterprise rollouts, or organizations that require independently audited security and measured speech-quality evidence before adoption.
local transcription and summarization by default; local storage; offline operation after model download; free core product; no account; AGPL source availability; read-only MCP access; system-audio meeting mode without inviting a bot; Notion/Slack/Calendar and file-export workflows; Keychain storage for secrets. 12
Apple-Silicon-only; several-gigabyte model footprint; early-stage real-world validation; no eligible external exact-product rating corpus; no independent accuracy or performance testing found; optional cloud/integration paths necessarily change the otherwise local data boundary. The current product page also references different local summary-model versions in separate sections, so exact bundled-model identity should be treated as changeable implementation detail rather than a stable product guarantee. 1
Products matched by shared tasks, audience, workflow, product type and capabilities.