How noncuro/muesli works
What this project is (type + purpose)
From the file tree and README, Muesli is currently a very early-stage (WIP) local desktop/tooling project (not a web app yet) intended to:
• Capture the last N seconds/minutes of audio from microphone and/or system audio
• Keep it locally in-memory with an auto-expiring buffer/queue (privacy principle)
• On-demand transcribe the recent audio snippet
• Use an LLM (gpt-4o) to rewrite/complete meeting notes
• Eventually provide an autocomplete UX (tab-to-accept) or shortcuts / popup insertion
• The target architecture (what the tool is aiming to become)
• Plus the current skeleton (a single Python entrypoint packaged as a tool)
Top-level files
• main.py: likely single entrypoint that will orchestrate capture → buffer → transcribe → LLM → UI insertion.
• setup.py: indicates intent to package/install as a Python tool (CLI and/or desktop entrypoint).
• .idea, .DS_Store: not architectural.
• No src/, no tests, no modules yet: architecture is conceptual right now.
Hotkey / Trigger Listener
• e.g., “capture last 30s–5m”, “transcribe now”, “insert completion”
Ring Buffer / Time-window Queue
• stores PCM/audio frames
• enforces retention policy (auto-delete > ~5m)
Buffer Manager / Retention Policy
• periodic cleanup
• exposes “get last X seconds”
Transcription Service Adapter
• “transcribe(audio_segment) -> text”
Transcript Cache (optional)
• short-lived in memory; may help with repeated prompts
Notes Input Source
• Later: “current application text context” (harder; depends on OS accessibility APIs)
Prompt Builder
• combines:
• existing notes so far
• transcript snippet
• user prompt templates (customizable principle)
LLM Client (gpt-4o)
• external dependency
• receives prompt, returns completion / rewrite
Completion Post-Processor
• formatting, trimming, safety checks
• handles partial suggestions vs full rewrite
Popup UI / Suggestion Panel
• shows suggestion, allows copy/select
Text Insertion Layer
• keyboard simulation / accessibility insertion into whatever app is active
Autocomplete UI (future)
• inline ghost text + tab accept (requires deeper integration)
Config/Prompt Templates
• local config file (not present yet, but planned)
Plugin/Adapter Interfaces
• transcription backend, LLM backend, UI mode, hotkey mappings
Local Logging + Debug Diagnostics
• Sensitive audio/text isn’t persisted unintentionally
Control / policy flows
• Retention Policy Timer → Ring Buffer
• Config/Prompts → Prompt Builder / Hotkeys / UI
Privacy-first ephemeral data design
• “in-memory ring buffer” is the core differentiator: show it prominently.
Adapter/ports-and-adapters (hexagonal)
• Audio capture, transcription backend, LLM backend, and UI insertion should be swappable.
Event-driven triggers
• hotkeys and timers trigger actions; audio capture is a stream.
Separation of concerns
• /buffering ≠ transcription ≠ LLM prompting ≠ UI/insertion
What technologies to label (given what’s known + what’s planned)
You only *know* it’s Python (via setup.py, main.py). The README mentions gpt-4o explicitly.
• Language/runtime: Python (local tool)
• LLM: OpenAI GPT-4o (external)
• Hotkeys/UI: “Hotkey Listener (TBD)”, “Popup UI (TBD)”, “Text insertion (OS accessibility / keyboard events) (TBD)”
This is still accurate: it shows intentional boundaries even if library choices aren’t made.