How it works

This guide is the technical picture of what happens when your AI answers from your company brain: how it connects, what it reads first, how it finds the right page, and the tools it has. If you want to know what is actually going on under a good answer, this is it.

How your AI reaches your brain

Each brain is its own private MCP server. Your AI client (Claude, ChatGPT, Gemini, or any MCP-capable client) connects to that server over MCP, authenticated as you. It sees exactly what you can see and nothing more. There is no separate app to open and no copy of the brain on your machine. The AI reaches into the live brain, filtered to your access, every time.

The session bootstrap

At the start of a session, the AI calls get_context_pack. This returns four core files, already filtered to what you are allowed to see:

company.md: what the company is.
brain.md: how this brain works and how the AI should use it.
index.md: a lightweight department router. It says where things live, not everything that exists. It is a map, never a full page list.
AGENTS.md: how the AI should behave in this brain.

Those four files orient the AI before it answers anything. They are small on purpose, so the AI always knows the shape of the brain without loading the whole thing.

Structured descent, not vector search

The AI does not search a pile of embeddings. It navigates on purpose:

Start from the core files.
Go to the right department hub, using index.md as the router.
Read the specific pages inside that department.
Open a source file only when a page cites it.

When a question is fuzzy or crosses departments, search_brain is the fallback. It is lexical search over the text, not vector or embedding similarity. So the normal path is structured descent through hubs and pages, with lexical search as a backstop. This is why answers can name the exact page they came from.

The tool surface

These are the tools the brain exposes to your AI. Each does one thing:

get_context_pack: fetch the four core files at the start of a session.
search_brain: lexical search across the pages you can see.
read_page: read one page in full.
walk_graph: follow the links between pages (people, projects, decisions) to see how they connect.
list_pages: browse what pages exist in an area.
list_changes: the authoritative, ordered history of what changed (created, edited, deleted), by whom and when.
list_skills / get_skill: discover saved procedures and fetch one to run.
list_systems / get_system: discover your connected tools and how to reach them.
ingest_to_brain: bring new material in and turn it into pages.

Every one of these respects your access. A tool never returns a page your role cannot see.

The Systems Registry

Your other tools (a CRM, a docs suite, a project tracker) each get a card in the brain, under wiki/tools/. list_systems lets the AI discover them and get_system tells it how to reach a given one, so it routes to that tool through your own connector instead of guessing. The cards describe the system and the route. They never hold credentials.

Skills

A skill is a saved procedure the AI runs on demand: a repeatable task written down once so it runs the same way every time. The AI finds them with list_skills and loads one with get_skill. Skills are turned on and off by status, so a team can keep a library and only expose the ones that should be live.

Where to go for more

Learning paths: where to start based on your role.
Best practices and what to avoid: how to keep the brain giving good answers.
The LAN docs (the source of truth): https://www.lanbrain.ai/docs