DEV Community

Nova Ops
Nova Ops

Posted on

No API Key, No Cloud, No Telemetry — Building an Obsidian Plugin That Works Offline

No API Key, No Cloud, No Telemetry — Building an Obsidian Plugin That Works Offline

Every "AI for your notes" tool we looked at wanted the same three things: an API key, a monthly bill, and your vault's contents flowing to somebody's server. So we built the opposite and shipped it: Elias HiveMind, a free, MIT-licensed Obsidian plugin whose core features make no network requests at all.

This is the design bet, and what it costs.

What "offline" actually has to cover

A plugin that only summarizes isn't enough — the vault chores are the real pile-up. HiveMind's local engine covers all of them with zero network:

  • Summaries — a local extractive summarizer: strip the markdown machinery, score sentences by meaningful-word frequency (length-normalized, early sentences slightly boosted), keep the best in original order. The same note always produces the same summary. Deterministic beats clever when a user is deciding whether to trust the output.
  • Tag suggestions — local, and vault-aware: it reads the tags your vault already uses (metadata only, never note content) and prefers them — casing, nested forms and all — over words it would otherwise invent.
  • Housekeeping sweep — the tagger run over a whole vault or folder, capped at 50 notes, as a dry run: one review list, nothing written until you confirm, note bodies never touched.
  • Link suggestions — finds notes whose title or alias appears in your text as a whole word but isn't linked yet, skipping frontmatter, code, and existing links.
  • Weekly + backlog digests — the week's notes summarized into one digest note, and the saved-articles pile (default folders: Clippings, Inbox) bucketed into Worth reading now / Reference / Gone stale / Junk candidates — each bucket saying why, in plain words. One run digests at most 50 clippings so a 500-article pile stays fast on a phone.

And the privacy line is written into the README, not the marketing: local commands make no network requests, the plugin collects and sends no usage data, and it never deletes anything or moves a note on its own.

The honest trade

An extractive summary is not a GPT summary. It rearranges your own sentences; it doesn't synthesize. For a digest you're skimming to decide what deserves your attention, that's usually the right tool — and it's free, instant, and private. When you genuinely want a model's rewrite, HiveMind can talk to a local Ollama server (default http://127.0.0.1:11434) or any OpenAI-compatible endpoint you configure — and if that server is unreachable, it falls back to the local summarizer instead of failing. Your notes still don't leave your machine unless you point the URL somewhere else yourself.

The constraint shaped the roadmap, too. A backend-powered weekly digest would cost one model call per note — so the digest is local-only by design. Limits you're honest about become features users can rely on.

Where it stands

HiveMind is live in the Obsidian community store (current release v1.3.0, shipped November 18) and the code is public at the link above. The backlog digest started as the whole product; the vault-aware tagger and the dry-run sweep grew out of using it on real, messy vaults — which is where every feature on the list came from.

If your notes are yours, the tool that reads them should act like it.


Facts in this post are from the project's public README and release record, checked 2026-11-20: feature set, the 50-note caps, the privacy statements, MIT license, and the v1.3.0 release. The plugin is free; there is nothing to buy in this post.

Top comments (0)