EN
Back to the archive

The archive · Developer & Business Tools · Product decision · 2026

Tolaria bets git-first markdown vaults are the memory AI agents work in

A solo dev's offline-first markdown knowledge app for humans and AI hit the HN front page and 10K GitHub stars within two weeks of launch.

Tolaria (RefactoringHQ)

The betThat knowledge bases should be git-first markdown vaults — offline, file-based, types as lenses, AI a first-class writer via MCP — not another plugin-locked notes app.Live

What the business is

Tolaria is a free, AGPL-licensed desktop app for macOS, Windows and Linux that manages markdown knowledge bases: notes with YAML frontmatter, types and relationships, git as the sync and version layer, an MCP server, and AGENTS files so Claude Code, Codex CLI and Gemini CLI treat a vault as work memory.

How it started

Luca, author of the Refactoring newsletter for more than six years, built Tolaria for himself to manage a workspace of 10,000+ notes and to test AI-assisted development on a real product, then launched it on Hacker News on 2026-04-23.

What happened

The Show HN drew 318 points and 142 comments, mostly engineers comparing Tolaria with Obsidian and Logseq; the AI-first positioning — git authorship, MCP, types — carried the debate. Run part-time alongside the newsletter, the project absorbed 200+ issues, 150+ PRs and 150+ feature requests in two weeks, which the author triaged down to 15 open issues and 81 feature requests.

How it ended up

Still live and actively developed as of 2026-09-05, with 3,700+ commits, a Homebrew cask and an open AGENTS-based workflow for AI contributors.

Background

Tolaria is a free, open-source (AGPL-3.0) desktop app for macOS, Windows and Linux that manages markdown knowledge bases. Notes are plain files with YAML frontmatter; types and relationships organize them; and every vault is a git repository, so history, remotes and version control come built in while the app adds no lock-in.

The bet is that AI changes what a notes app should be: an agent that reads and writes files needs a git-first, file-first workspace it can act on. Tolaria therefore ships an MCP server and AGENTS files for Claude Code, Codex CLI and Gemini CLI rather than a plugin ecosystem.

Luca, author of the Refactoring newsletter, built Tolaria for his own 10,000+ note workspace and launched it on Hacker News on 2026-04-23. The Show HN drew 318 points and 142 comments, and within two weeks the repo passed 10K stars and several thousand daily users while absorbing 200+ issues and 150+ PRs.

As of 2026-09-05 the project remains live and is developed part-time: 3,700+ commits, a Homebrew cask, and a maintenance workflow that keeps open issues in the tens while a community of contributors ships fixes.

What has to be true

  • Obsidian and Logseq keep markdown local too, but they are built around a human curator; Tolaria made the AI agent a first-class writer with git authorship, MCP and an AGENTS file.
  • Git-first design gave users full history and any remote for free, removing a sync server from the trust equation entirely.
  • Plain files plus YAML frontmatter meant zero lock-in, so skeptical engineers could adopt Tolaria without committing their data to the app.
  • The author's own 10,000-note vault was real dogfooding: every feature had to solve a problem he actually had.

What can be applied

Open-source distribution can outrun any go-to-market plan: a free, file-first tool built for one writer's 10,000-note vault reached several thousand daily users and 10K stars in two weeks.

Aftermath

As of 2026-09-05 Tolaria is still live and actively developed by Luca part-time while he runs the Refactoring newsletter. The repo stands at 3,700+ commits with a Homebrew cask and releases for macOS, Windows and Linux, and the maintenance workflow he published in May 2026 — bugs on GitHub Issues, feature requests on Canny, AI agents triaging both — is part of the product's story. No funding round or company formation has been reported in the sources reviewed; the bet remains that a free, open, file-first workspace is the memory layer AI agents work in.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases