michaelwilhelmsen/humla

What it solves

Humla is a local-first meeting notes application for macOS that allows users to capture freeform notes during meetings while simultaneously recording and transcribing audio. It solves the problem of manually summarizing meetings by combining user-written notes with an automated transcript to produce structured summaries, all while ensuring data privacy by keeping audio and text on the device.

How it works

The app captures microphone and system audio as two separate streams to ensure clean transcription and speaker separation. It supports multiple transcription engines, including local Whisper models (optimized for Apple Silicon) and cloud providers like OpenAI, Deepgram, and Groq. After recording, it runs an offline speaker-identification pass using models like Nemotron 3 or Community-1. For summarization and chat, it integrates with OpenAI or local LLM servers (Ollama, LM Studio, llama.cpp). It also includes a Model Context Protocol (MCP) server that allows external AI agents (like Claude Code) to search and read meeting notes.

Who it’s for

Professionals, students, and researchers on macOS who need accurate, private meeting records and summaries without relying on on-device bots or cloud-only SaaS tools.

Highlights

• Local-First Privacy: Audio and data stay on the machine by default; supports fully on-device transcription, speaker identification, and summarization.
• Dual-Stream Recording: Captures mic and system audio separately to prevent voice mixing and improve transcript clarity.
• Flexible Transcription: Mix-and-match providers per language, including a specialized Norwegian-tuned Whisper model.
• Hybrid Search & Chat: Combines full-text and semantic embedding search to answer questions about meeting history with citations.
• MCP Integration: Ships with a read-only MCP server for AI agents to query meeting notes directly.
• Self-Hostable Sync: Optional team synchronization via a PocketBase backend that can be self-hosted for free.