Timaga/noda-local-ai-notes

What it is

Noda is a Windows desktop note-taking app that keeps each project's notes as Markdown files in a folder you choose, with app settings and chat history stored separately in the Windows profile. It adds optional local AI stages on top of the notes: generated titles and tags, links and saved connections, a knowledge graph, semantic search, and chat that quotes and links the source notes it retrieved. Everything runs against models installed on the machine; there is no hosted model API in the normal workflow.

The target user is someone who already writes Markdown and wants retrieval and cross-note questions without sending the library to a cloud service. The graph shows saved relationships rather than inferred topical similarity, and AI-generated metadata stays editable.

Install and first run

The supported path is the Windows x64 setup from the releases page. Download Noda-Setup-<version>-x64.exe, run it, and launch Noda from the Start menu. The installer is unsigned while code-signing is pending; the project asks you to read the known security issues before installing.

On first run, create a project or connect an existing Markdown folder. Setup downloads the semantic-search model (about 1.1 GB) and then lets you choose processing models or disable optional AI stages; model weights are not bundled. Create a quick note, confirm it appears in the processing queue, and open the result in the library.

Automatic app updates are not included: back up projects and profile, exit Noda fully from the tray menu, then run the newer setup. Project folders and profile data survive reinstalls.

Build from source

Developers need Windows x64, Git, and Node.js 22.12 or newer within the 22.x line with npm. In PowerShell:

git clone https://github.com/Timaga/noda-local-ai-notes.git
cd noda-local-ai-notes
npm --prefix frontend ci
npm --prefix frontend run build

The build script prints the generated Setup path and does not publish it. Tests are described in the development guide, and version preparation in the release guide.

Requirements and limits

• Windows 10 or 11 x64 on an AVX2-capable CPU; 16 GB RAM is described as a useful starting point for the recommended model profile, and larger models need more.
• Browsing, editing and plain text search work without the semantic-search model; semantic search and note chat need their respective local models. CPU processing can take minutes.
• Supplied note-processing profiles support Russian and English only. The interface has ten languages, but changing interface language does not translate notes.
• Noda manages local Python and Qdrant components and downloads runtime dependencies during setup, so initial setup needs internet access and extra disk space.
• Project files are not encrypted; the README points to OS-level storage protection and backups. Fine-tuning is disabled, and exact-binary clean-machine validation and broader model evaluation remain documented gaps.
• The code is under PolyForm Noncommercial License 1.0.0 (source-available, not OSI open source); commercial use is not granted. Third-party runtimes, models and assets keep their own licenses.

Chat answers and generated metadata can be wrong; the README advises checking cited text when accuracy matters. Screenshots use synthetic demo notes.

HotRepo's take

The only station record is a short new-list snapshot on 2026-11-15 (JavaScript list), where the project gained 3 stars over roughly one hour, reaching 105. That is a single observation, too thin to describe a trend. The notable facts for readers are the deliberately narrow setup: Markdown-owned data, cited retrieval, and a noncommercial source-available license, combined with an unsigned installer whose binary-to-source relationship and real-world usage are not established by the repository materials.