Obsidian and AI: building the knowledge base your agent reads

Obsidian is a note-taking app that saves everything as Markdown files, inside an ordinary folder on your computer. No hidden database, no proprietary format. That apparent dullness became its greatest asset the day artificial intelligence agents learned to open files. A well-kept folder of notes is a company memory an AI can read, enrich and keep up to date.

This tutorial is written for a founder or a team lead. It describes the four-step setup, in order. It also flags the three points that cost real money when you discover them afterwards, and that no French tutorial mentions.

What Obsidian is, and why AI gave it a second wind

Obsidian opens a folder and displays the Markdown files inside it. You write, you link notes together, the app draws the graph of those links. The files stay readable in any text editor, today and in ten years.

The vendor's security page is explicit about the rest. Data is saved locally, and the page states that no account is required and no telemetry data is collected. The desktop and mobile clients were audited by Cure53 three times between December 2023 and December 2024, and the sync service by Trail of Bits in December 2025. The vendor published the results of the last two audits on its blog in May 2026.

What changed is the reader. For years those notes served one person. Now that Claude, ChatGPT or a homemade agent can walk through a folder, the same vault feeds context to a machine. The question many companies ask therefore shifts. It becomes: where do we keep what the company knows, so an AI can use it?

I had sorted the tools by use case in my pyramid of AI uses published in January 2026. The research and knowledge category gathered three hosted services. The gap is obvious a year later. There was no slot for the knowledge that has to stay on your own machine.

Three ways to wire AI to your notes

AI does not ship with the app. Obsidian provides no built-in assistant, and that is a deliberate choice by the vendor. You therefore add the intelligence on top, in three ways that combine well.

1. A community plugin. You install a module inside the app. It reads your vault and talks to a model, usually with your own API key.
2. An agent wired to the folder. Claude Code or an equivalent agent opens the files the way a colleague would. You name the folder, the rest of the disk stays out of reach.
3. A model installed on the machine. Nothing leaves the workstation, at the cost of lower quality and a machine able to run it.

The third route deserves its own article, and we wrote a full guide to running AI locally in a company. The rest of this tutorial takes the second one, which offers the best ratio between setup effort and control gained.

Step 1, create the vault and separate sources from the wiki

An Obsidian vault is a folder. You create it in three clicks, and the real decision comes right after. It concerns the internal layout, and it determines whether your base survives six months.

The architecture that holds separates three folders with different write rights. A sources folder where you drop raw material, meeting notes, exports, received documents, and which the agent reads without ever touching it. A wiki folder where the agent writes and updates the entries. A rules file that you write and it must obey.

That separation comes from our own practice. We documented it in the guide to building a second brain with Obsidian and Claude Code, built over a weekend and used internally ever since. Allow around forty-five minutes to set it up.

Step 2, give your notes properties so the AI can navigate

A note without labels forces the agent to re-read everything before answering. Obsidian solves this with properties, a small block at the top of each file carrying structured fields. The official documentation lists seven types, from plain text to date and time, including numbers and checkboxes.

Three or four properties are enough at the start. The note type, the client or project involved, the date, the status. The agent uses them to filter before reading, which lowers the cost of each query and the number of mistakes.

A core plugin called Bases then turns those properties into table, list, card or kanban views. No separate database is created. Everything stays inside your Markdown files, so everything stays readable by an agent. The logic matches the one described in our guide to RAG and business context, applied to files rather than to a vector store.

Step 3, wire your agent to the folder

Two routes exist, and both are official. The first goes through the command line published by the vendor. Its help page states that anything you can do in Obsidian you can do from the command line, and explicitly mentions giving agentic tools access. The app must be running and the feature must be enabled in settings.

The second route goes through the MCP protocol, published by Anthropic in November 2024. An MCP server exposes your vault to the agent as a set of tools: search a note, create one, edit one. Several community servers exist for Obsidian. Some read the files directly, others talk to the running app.

In that video I show an agent wired to a hosted note-taking tool. Every action had to be declared one by one: create a page, update a property, retrieve a block, then maintain those declarations each time the documentation changed. An MCP server replaces that work with a single connection, kept up to date by whoever publishes it.

Step 4, write the rules the agent must follow

The last step is the one almost everyone skips, and it makes the difference between a living base and a pile of files. You write an instructions file at the root of the vault. The agent reads it before every intervention.

• The format of entries. A client entry contains these sections, in this order, with these properties.
• The house vocabulary. The words your company uses, and the ones it never uses.
• The forbidden folders. What the agent reads, what it edits, what it must ignore entirely.
• What never gets stored. Sensitive personal data, passwords, contractual documents that have not been anonymised.

This file plays the role that I described in my newsletter issue on context assigned to RAG for a business agent. A model without context stays mediocre, and a model with clear rules becomes reliable. The difference rarely comes down to model quality.

The plugin trap: nothing is sandboxed

Here is the point French tutorials leave out. Obsidian's community catalogue is public, and it reads straight from the vendor's official repository. In October 2026 it holds more than 8,800 published plugins. Roughly one in ten advertises an AI feature in its description.

The official page on plugin security is blunt. It states that due to technical limitations, Obsidian cannot reliably restrict plugins to specific permissions. An installed plugin can therefore access files on your computer, connect to the internet and install additional programs. The vendor automatically scans every version, reserves manual reviews for popular, featured or flagged plugins, and publishes the result as a safety scorecard. Its recommendation for sensitive data fits in one line: have the plugin audited before using it.

That table does two jobs. It gives the only objective criterion for picking a plugin in a tool that sandboxes nothing: the install count, and therefore the number of eyes on the code. And it restates a useful obvious fact. The most installed plugins in the catalogue are for drawing diagrams and automating note templates. AI remains an added layer, useful and in the minority.

In a company, Tandem's rule is simple. Restricted mode stays on by default, and every enabled plugin is an explicit decision. It is the same discipline described in our article on AI agent security in business. A plugin holds the rights of whoever installs it.

Your notes sit in the clear, and sync has a region

Second rarely covered point. Obsidian never encrypts your local vault. Its documentation says so in one sentence. On a laptop, your operating system's disk encryption therefore remains the only protection in case of theft. That setting takes two minutes to check and it concerns the whole company, well beyond notes.

The vendor's sync service runs on DigitalOcean, in Singapore, Frankfurt, San Francisco and Sydney. Frankfurt is the only European Union option. The automatic setting picks a region from your IP address at first setup, which exposes a team configured from another continent. And the documentation warns that changing region means recreating the remote vault, losing version history along the way. The decision is therefore made on day one.

What Obsidian costs in a company

The app is free, with no limit on notes or vaults. Three paid options exist, and each is decided separately.

• The commercial licence. About forty euros per person per year, for 50 dollars listed in October 2026. The pricing page states in plain words that you are not required to pay it.
• Sync. A little under 4 euros per person per month on an annual plan, for 4 dollars listed. It is the only genuinely useful line for a team.
• Web publishing. Around 7 euros per site per month on an annual plan, for 8 dollars listed. Useful for public documentation, pointless for an internal vault.

The counterpoint is enjoyable, and it checks out on the vendor's own site. The licence stays optional, and more than 10,000 organisations report paying it anyway. The list is public, sorted by support tier. On the European side you will read Thales, Siemens, Bosch, BASF, Deutsche Telekom, Fraunhofer, TÜV, DATEV and the UK Government. For a hesitant management team, that list beats any sales pitch.

Per workstation, the annual budget lands around 90 euros per person, licence and sync included. Compare that with the seats of a hosted workspace, and with the cost of the meeting tools we reviewed in our comparison of AI note-taking tools.

Should you install Obsidian to work with AI?

Yes in one precise case, and no in the others. Obsidian earns its place when your knowledge has to stay on your machines, when several agents must read the same base, and when you accept the filing discipline that comes with it. Those three conditions travel together.

If your need stops at questioning a few documents without keeping them, a hosted tool will be faster. Our tutorial on NotebookLM, now Gemini Notebook covers that case, and it requires zero installation. If instead you want knowledge to accumulate rather than be rebuilt at every question, the folder of files wins.

Across around a hundred engagements, the obstacle Tandem meets most often has nothing to do with tooling. Company knowledge lives in chat threads, mail chains and people's heads. When a key colleague leaves, the knowledge of a project leaves with them. A vault of AI-readable files does not solve that on its own. It gives it somewhere to exist.

The real work starts afterwards, and it is human. Deciding what gets in, what stays out, who reviews. That is the heart of our AI enablement programmes, and no plugin will do it for you.