Every practice I know has a knowledge base that lives in three places at once: a SharePoint nobody trusts, a Notion nobody pays for anymore, and the one Bauleiter’s head. The files that actually matter — the detail notes, the BEP fragments, the “why we did the facade like that” memo — are almost always plain text somewhere. The tool for editing them keeps changing. The text does not.

That is the quiet argument behind OpenKnowledge, posted to Hacker News this month by the team at Inkeep as a GNU GPL v3.0-or-later alternative to Obsidian and Notion. It edits markdown and MDX files sitting in a folder on your disk, with a true WYSIWYG surface that feels like a Google Doc, and it wires AI agents — Claude Code, Claude Desktop, Cursor, Codex, OpenCode, Pi — directly at the same folder over MCP. Notion meets VS Code, as their README puts it, which is a fair description and also an admission about who it is for.

←TODAY: Your project knowledge is markdown in a folder; the editor is rented, the folder is yours.
→3012: The archives that survived to Zurich-3012 were the plain-text ones — nobody wrote a reader for a dead note app.
Fulcrum: An AI that can read your office’s knowledge is only useful if the knowledge outlives the AI’s vendor.

The Tool:

OpenKnowledge (inkeep/open-knowledge on GitHub) is a local-first markdown and MDX editor with a file navigator, search, tabs, and a graph view for wiki links. It ships as a macOS DMG and, for everyone else — Linux, Windows, Intel Mac — as a Node.js 24+ CLI that serves the same editor as a local web app. Sync and team sharing run on git and GitHub under the hood — no proprietary sync server, no seat licence. The part worth an architect’s afternoon is the MCP layer: ok init installs skills and an agentic search endpoint for every harness it detects on your machine, so a coding agent can read and author documents in your vault without you pasting anything into a chat box. It also renders embeddable HTML and rich components, which is the feature that makes engineering specs and visualised reports plausible rather than painful.

Because your office already has the corpus. It is just unreadable to anything except a human with time.

Setup:

# Node.js 24+ and git required
npm install -g @inkeep/open-knowledge

# point it at real content, not an empty folder
cd ~/projects/office-wiki
git init          # skip if already a repo

ok init           # scaffolds + wires detected AI editors (Claude Code, Cursor, Codex...)
ok start --open   # serves the editor, opens your browser

On macOS you can skip the CLI entirely: download the DMG from the latest release, drag it to Applications, launch. Either way, you open an existing folder — an old Obsidian vault, a docs directory inside a codebase, the pile of .md files you exported from Notion in a fit of principle.

First steps:

1. Open a folder that already hurts. Not a demo vault. Take the directory where your standard-detail notes or your BEP drafts live. If they are in .docx, run pandoc -t gfm over them first — the conversion loss is real but survivable, and it is one-way out of a format you do not control.
2. Make two wiki links. In any note, write [[IFC-Export-Settings]] and [[Fassade-U-Wert-Nachweis]]. Open the graph view. You now have the smallest possible picture of how your office’s knowledge actually connects — usually a shock, usually two orphan clusters and one hub.
3. Let the agent use the MCP wiring ok init installed. Then ask your harness a question you would normally ask a colleague: “which projects used the same window detail as the one in 2024-notes/?” The answer quality tells you more about your note hygiene than about the model.
4. Commit. git add -A && git commit -m "wiki: first pass". Sync is just git. That is the whole feature, and it is the right amount of feature.

The part where I get sceptical

Let me be plain about the trade-off, because the app-roundup press will not be. The XDA piece on Obsidian alternatives makes the honest point that these tools complement rather than replace — and moving an office to a new knowledge editor has a real cost in retraining and half-migrated content. OpenKnowledge is young. WYSIWYG markdown editors have a long history of quietly mangling tables, footnotes, and anything with embedded HTML. Test it on a copy of your repo, not the repo.

But the risk is bounded in a way that matters, and this is the whole reason I am writing about a note app on a BIM desk. If OpenKnowledge disappears next year, you still have a folder of markdown in a git repo. Every editor that has ever existed can open it. Compare that to the failure mode of a hosted block-based workspace, where “export” means a ZIP of HTML with generated filenames and your links rot on the way out.

Why this is a BIM problem, not a stationery problem

The How-To Geek roundup of overlooked open-source tools frames apps like this as personal-productivity finds. For a practice, it is closer to a records question. Our industry has thirty years of evidence about what happens to project knowledge stored in a format tied to a company: it becomes archaeology. I spend a meaningful share of my working life on IFC round-trips precisely because someone, somewhere, decided the native format was good enough forever.

The buildings that aged badly were never the ugly ones. They were the ones whose documentation nobody could open. If you are choosing a knowledge stack this quarter, ask one question: when the vendor goes dark, can a 25-year-old still read the file? Markdown in git answers yes. Almost nothing else does.

There is a second-order effect too. A plain-text, git-backed vault is the cheapest possible substrate for the agentic search everyone is now trying to buy as a service. You do not need a RAG vendor to index your office wiki if your office wiki is already 400 text files with sane names. PAZ’s own Creative Coding — En Ingenieria panel argues the expressive sketch and the production script are the same skill at two altitudes; the same logic runs here. The file format you can grep is the file format your agent can reason over.

Atelier: For an office adopting AI this year, the binding constraint is not model quality — it is that your institutional knowledge is locked in PDFs, emails, and a Notion workspace one person administers. OpenKnowledge is a cheap way to test whether a git-backed markdown vault plus an MCP-connected agent can actually answer real project questions, before you sign anything. Monday move: pick your single most-referenced internal document — the office IFC export checklist, the standard-detail index, whatever people screenshot to each other — convert it to markdown, put it in a git repo, and run ok start on it. One file. If the agent can answer questions about that one file, you have your pilot.

Hack: Find out which of your notes are orphans before you trust the graph view. Wiki links only help if they exist, and most migrated vaults are 80% unlinked debris. Run this from the root of your markdown folder and it lists every .md file that nothing else links to — your dead ends, ranked by shame. Fix the top ten by hand, then re-run.

import re, pathlib
files = {p.stem: p for p in pathlib.Path(".").rglob("*.md")}
linked = set(re.findall(r"\[\[([^\]|#]+)", "\n".join(p.read_text(errors="ignore") for p in files.values())))
for stem in sorted(set(files) - {l.strip() for l in linked}):
    print(stem)

Learn-it:

• The repo: github.com/inkeep/open-knowledge — GPL v3.0-or-later, issues and PRs open.
• Docs: openknowledge.ai/docs — general usage, starter packs for LLM wikis and second brains.
• Context on the alternatives: XDA — when your Obsidian vault gets too bloated, a useful corrective on migration cost.
• The wider pattern: How-To Geek — overlooked open-source apps.
• PAZ note: the same argument runs through our IFC and format-longevity coverage — the point of open formats was never ideology, it was being able to open the file in 2051.

So: clone it, point it at one folder that already causes arguments in your office, and see whether the agent can answer a question you are tired of answering. Run it on a copy, on your ugliest directory, and let the graph view tell you what your office actually knows.