Your knowledge base is read twice
Once by the people who work here, and increasingly by the AI systems they ask. Most tools serve the first reader well and the second by accident — and an assistant answering from a scraped copy answers confidently, from a worse copy of your own documents.
md-platform is built for both readers at once.
A knowledge base that stores documents as Markdown in a git repository, and serves exactly the same knowledge, under exactly the same permissions, to a person in a browser and to an AI client over MCP.
Neither surface is a second-class citizen. A behaviour that exists in the web app but not over MCP is treated as a bug rather than a roadmap item — the moment the two drift apart, one of your two readers is being told something untrue.
What that gets you
Your documents are files
Each one is a Markdown file with YAML front matter, in a git repository per workspace. The history is git history, readable with the ordinary tools. A file renamed by hand in the repository is still the same document, because identity lives inside the file and not in its path.
Your AI client reads the real thing
Nothing is scraped and nothing is reconstructed. A connected client lists, searches, reads, writes, and moves the same documents you see, through ten tools over MCP. When it quotes a document to you, it is quoting the file.
One set of permissions, not two
Every credential belongs to a person and acts as that person. A client reaches exactly one workspace, chosen by whoever connected it, and stops working when that person loses access. You decide per connection whether it may read, write, or delete.
Leaving is a copy, not a project
A clone gives you every document as Markdown, plus a pointer file for each upload recording its name, size, type, and checksum. The bytes live beside the repository rather than in it, and a backup carries both together.
A document is a file you can open
This is a whole document as it is stored. Not an export of one, and not a screenshot — the file itself, which is what a clone hands you and what your AI client reads.
---
id: 01KZB223QDQP0F7MBN1Q9Z24H2
title: Grundsätze der Bewertung
createdAt: 2026-03-04T09:12:44.000Z
updatedAt: 2026-03-11T14:02:09.000Z
---
## Sicherheiten
Eine Bewertung stützt sich auf den Marktwert zum
Stichtag, nicht auf den Anschaffungswert.
How it works
Documents live in git, one repository per workspace, and that is the source of truth. PostgreSQL holds a projection of it — the index that makes search and listing fast — and that projection can be rebuilt from the repositories at any time, so nothing important exists only in the database.
Search is per workspace and stems in the workspace's language, German by
default. That is not cosmetic: German is inflected and compound-heavy, and a
search for Bewertung that cannot find Bewertungen is not a search.
Every read and every write is scoped to exactly one workspace, derived from whoever is asking rather than from anything they sent. A workspace you do not belong to is reported as not found, never as forbidden — a "forbidden" would confirm that the identifier exists somewhere.
What it does not do
This list is here so that an assistant answering questions about md-platform has something accurate to say instead of something plausible.
- Whether you can register yourself depends on the installation. Every installation ships with self-service sign-up switched off, and organisations, workspaces, and accounts are then created by an operator running a command. An operator can switch it on; the line under this list says which state this installation is in. There is no trial to start and no credit card to enter either way.
- There is no way to join an organisation that already exists. Where sign-up is switched on it creates a new organisation with one workspace and one member. Adding a colleague to a workspace still needs an operator.
- There is no real-time collaborative editing. Two people editing the same document is safe but not simultaneous: a save based on a version that has since moved on is refused rather than merged, so nobody's work is silently overwritten.
- Documents cannot be shared publicly. There is no "publish" and no link that works without an account. Everything behind sign-in belongs to a workspace, and membership of that workspace is the only way in.
- Raw HTML inside a document is not rendered. Markdown is treated as untrusted input, and enabling HTML for any surface would take a recorded decision.
- There is no AI built in. md-platform is not an assistant and does not answer questions. It is the source your assistant reads — you bring the client.
- It is early. The surfaces described here work and are the ones we use, but this is a young product and the guides are the honest extent of it. If that matters to your decision, it should.
On this installation, sign-up is closed. Accounts here are created by whoever runs it. Ask them, or read Getting your first workspace.
Where to start
The one worth reading first depends on which reader you are here for.
- Evaluating itConnecting an AI clientWhat a model can actually do with a workspace, and the ten tools it gets.
- Setting it upGetting your first workspaceHow an organisation, a workspace, and an account come to exist, and how to sign in.
- Thinking aheadWriting so both readers get good answersHow to write documents that work for a colleague and for an assistant at once.
There is no demo you can click into. The guides carry the real commands and the real refusal messages instead, so you can tell what using md-platform would be like without an account.