A developer named Nimendra just shipped NoteBrain, a local-first command-line tool that turns your Obsidian vault into a searchable knowledge backend for AI coding agents. The launch showed up as a Show HN post on Hacker News, where it climbed to 162 points. What stands out here is the approach: everything runs offline, on your own machine, with no cloud service touching your notes.
The pitch is simple. If you keep your second brain in Obsidian, NoteBrain lets an AI agent actually read and reason over it. It indexes your markdown notes into a local ChromaDB vector database, then exposes semantic search, wikilink graph traversal, and hidden-connection discovery through structured output that agents can chain directly into their workflows.
What NoteBrain Actually Does
- Semantic search that ignores keywords. It finds notes by meaning using the offline all-MiniLM-L6-v2 ONNX embedding model, so you don’t need exact-match terms to surface the right note.
- Knowledge graph traversal. It walks your Obsidian wikilink graph through backlinks, multi-hop connections, and shared tags, which mirrors how you actually think across notes.
- Hidden connections. This is the interesting one. It surfaces semantically related notes that you never explicitly linked, with an optional deep mode that matches section by section.
- Graph-boosted ranking. It blends semantic similarity with graph relationships to sharpen relevance, rather than relying on vector math alone.
- Structured output for automation. Results export as JSON or TSV with built-in JSONPath querying, so you can pipe them straight into shell scripts or agent tool-use loops.
Built for Agents, Not Just Humans
The reason this matters goes beyond note search. NoteBrain ships with an AI agent skill and an OpenCode agent configuration, and it’s designed to plug into autonomous coding agents like OpenCode, Google Antigravity, Pi Agent, and Claude Code. According to the Hacker News post, the setup is specifically tuned to cut token usage and latency, two costs that add up fast when an agent queries your knowledge base repeatedly.
Nimendra recommends pairing it with the Pi Agent and the provided skill, noting it delivers solid results even on low-cost models like DeepSeek Flash without burning tokens, partly by improving cache hit rates.
There’s real engineering under the hood. NoteBrain uses Goldmark AST-aware chunking, which splits markdown by header hierarchy instead of arbitrary character counts. That preserves lists, tables, callouts, and code blocks intact, so an agent gets clean context instead of fragments. Incremental ingestion with SHA-256 hashing skips unmodified notes in milliseconds on re-runs, so keeping your index fresh via a cron job stays cheap.
Availability and Caveats
NoteBrain is open source and free on GitHub. That said, there are real limits worth knowing before you commit.
- Linux-first. The developer built it for personal use on Linux. macOS and Windows binaries aren’t shipped and are listed as untested. You can compile from source, but you’re on your own for those platforms.
- Text only. It handles markdown and nothing else right now. No PDF parsing, no image OCR.
- Some assembly required. It needs Go 1.26.4+ with a CGO-enabled toolchain, and first-time indexing can take several minutes on a large vault.
Nimendra is refreshingly honest about the scope, calling it a personal project and openly encouraging people to fork and adapt it rather than treat it as a polished product.
Why It Matters
NoteBrain sits at a useful intersection. Plenty of people have years of notes locked in Obsidian, and plenty of people now want their AI agents to use that knowledge. Most existing solutions push your data to a cloud RAG service. This one keeps everything on disk, which is a genuine advantage for anyone who treats their notes as private.
The bigger signal is where agent tooling is heading. Local, composable, shell-friendly building blocks that agents can chain together are becoming their own category. Expect more tools that treat your personal data as an offline backend rather than something to upload. You can find the full details and source at the original Hacker News post.