Muesli 🥣

A privacy-focused, self-hosted meeting-notes app: self-hosted audio transcription and self-hosted LLM summarization, with your data staying on infrastructure you control.

Muesli is an open, self-hostable alternative to cloud meeting-notes apps. You record a meeting, and Muesli transcribes it and rewrites your sparse notes into a clean summary — but the transcription and summarization run as plugins you host (by default a self-hosted Whisper transcriber and a self-hosted Ollama LLM), so audio and notes never have to leave infrastructure you control.

Screenshots

⚠️ Placeholder — real screenshot pending. To replace: run npm run dev, open a note with a summarized body, take a screenshot (1280×800), save it as docs/images/notes-ui.png, then update the notes-ui.svg reference in this README to notes-ui.png.

Requirements

Before running docker compose up, make sure your host meets these prerequisites:

• Docker ≥ 24 and Compose v2 — the stack uses docker compose (plugin); the legacy docker-compose v1 standalone is not supported.
• RAM — ≥ 8 GB system RAM is recommended for comfortable CPU inference with the default llama3.2:3b model.
• Disk space — ~2 GB for the default LLM model; allow more if you plan to pull larger models. Whisper also fetches its model on the first transcription.
• CPU vs GPU — CPU inference works out of the box but is slow (the default dev setup); GPU acceleration is available — see the per-plugin Dockerfiles.

Why Muesli

• Private by default — the reference plugins run on your self-hosted server; nothing is shipped to a third-party transcription or LLM SaaS.
• Self-hostable — one docker compose up on a NAS or a cloud box. Scale the pieces independently when you need to.
• Pluggable — transcription and summarization are language-agnostic HTTP plugins. Swap in any engine that speaks the contract; bring your own cloud LLM if you choose to.
• Current v3 feature set — calendar integration with CalDAV / ICS / Google / Microsoft sync, note-linked events and Coming up; Chat / RAG over notes with citations; semantic search; speaker diarization with review and speaker aliases; export to Markdown, plain text, DOCX, PDF, and subtitle formats (SRT / ASS / WebVTT); tags, smart lists, folders, nesting, sibling reorder, and recycle bin / soft delete for notes and folders.
• Open source — AGPL-3.0; no lock-in, inspect everything.

Quickstart

Bring up the entire stack (Postgres, Ollama, Whisper transcriber, LLM agent, and the Muesli server) with one command:

docker compose up

Then open http://localhost:8080/admin, create your account, and you're done — the default transcriber and agent plugins are auto-registered on startup, so there are no manual admin steps.

First boot is slow: the ollama-pull service downloads the default LLM (llama3.2:3b, ~2 GB) and Whisper fetches its model on the first transcription. Expect a few minutes before the stack is fully ready. CPU inference is slow — this is a local-first dev setup, not a tuned production deployment. (For GPU acceleration, use the override described in the GPU acceleration section below.)

For a step-by-step first run, see docs/GETTING_STARTED.md.

Deploying to production? See docs/DEPLOYMENT.md.

One-line install

Use this for a production/hosted deployment. It downloads the GitHub Release asset bundle (a version-pinned docker-compose.prod.yml, .env.example, install.sh, and a SHA256SUMS file), verifies checksums before installing anything, generates real secrets in .env, and uses the GHCR-hosted images already pinned into that release's compose file.

curl -fsSL https://github.com/abedegno/muesli/releases/latest/download/install.sh | sh

This is separate from the dev docker compose up quickstart above. The installer writes the production files into ./muesli by default; set MUESLI_RELEASE_TAG to pin a specific release instead of the latest one, and use --up if you want the script to start the stack after installation.

GPU acceleration

An optional Compose override adds NVIDIA GPU support for Ollama and Whisper. Requires the NVIDIA Container Toolkit on the host.

docker compose -f docker-compose.yml -f docker-compose.gpu.yml up

The override sets runtime: nvidia + device reservations on the ollama service and switches Whisper to WHISPER_DEVICE=cuda / WHISPER_COMPUTE_TYPE=float16. All other services are unchanged.

No .env is required: every secret and URL has a built-in dev default. To override anything, cp .env.example .env and edit.

Troubleshooting: See docs/TROUBLESHOOTING.md for help with slow first boot, port conflicts, and other common issues.

Desktop client

The Electron desktop client lives at the repo root:

npm install
npm run dev

Point it at your running server: http://localhost:8080.

Native iOS client

A read-only, native SwiftUI iOS app lives at native/ios (issue #767): browse your existing notes and read their authored body and generated summaries, over the same hosted server or from a paired local Electron instance ("Allow iOS access" in Settings). Note creation, audio capture, and Android are explicitly out of scope for this first slice. See native/ios/README.md for project setup, the mobile API contract it consumes, and this repo's local-pairing security model.

How it works

Desktop client ──upload──▶ Muesli server ──▶ Transcriber plugin ──▶ Agent plugin ──▶ note
   (capture)              (queue + pipeline)     (Whisper)             (Ollama)

The server stores your notes and audio, runs a worker pool that drives each recording through transcribe → summarize, and serves an embedded admin UI at /admin. Plugins are independent HTTP services. For the full picture — package map, the plugin contract, and the processing pipeline — see docs/ARCHITECTURE.md.

Documentation

Doc	What's in it
docs/index.md	Master documentation index — all docs by use case.
docs/ARCHITECTURE.md	Contributor-oriented architecture: components, pipeline, package map, plugin contract, and the design rationale behind the privacy-first, self-hosted plugin approach.
docs/CONFIGURATION.md	Complete reference for all MUESLI_* and plugin environment variables.
docs/DEPLOYMENT.md	Production runbook: install, TLS, sizing, backups, upgrades, and smoke checks.
docs/API.md	HTTP API reference: JSON models, endpoints, auth rules, and response shapes.
docs/PLUGINS.md	Plugin authoring guide: the transcriber/agent contract, required endpoints, auth, and validation flow.
docs/BACKUP.md	Backup and restore guide for the Postgres database and audio blob store.
docs/UPGRADING.md	Upgrade procedure for pulling or rebuilding images, restarting, and handling migrations.
docs/TROUBLESHOOTING.md	Common issues: slow first boot, port conflicts, reading logs, and telling "still loading" from broken.
CONTRIBUTING.md	Dev setup, conventions, and how to land a change.
SECURITY.md	How to report a vulnerability (privately) + operator responsibilities.
CODE_OF_CONDUCT.md	Community standards.
CHANGELOG.md	What's changed.
ROADMAP.md	High-level milestones (v2 → v5) closing the gap to Granola and into enterprise.

Security note

The compose file ships DEV-ONLY default secrets (master key, storage signing key, plugin tokens, Postgres password) so docker compose up just works. Do not run these defaults in production. Before any real deployment, copy .env.example to .env and set real values — in particular generate a fresh MUESLI_MASTER_KEY and MUESLI_STORAGE_SIGNING_KEY with openssl rand -base64 32, and terminate TLS at a reverse proxy. See SECURITY.md.

Status

Muesli is pre-1.0 and under active development. The current v3 feature set covers calendar integration, Chat/RAG over notes, speaker diarization with review and speaker aliases, semantic search, multi-format export, and tags, smart lists, folders, and recycle bin support. The product is still evolving, and there is more polish and broader multi-user work ahead on the roadmap.

Contributing

Contributions are welcome — see CONTRIBUTING.md to get started, and please open an issue before tackling anything non-trivial.

License

GNU Affero General Public License v3.0. If you run a modified Muesli as a network service, the AGPL requires you to make your changes available to its users.