The model on your desk, the files on your disk, and nothing in between

Anybody who has gone to the trouble of running a model on their own hardware has usually done it for a reason that survives scrutiny: work under an NDA, a client’s material, a case, an investigation, or an employer whose policy is simply no. Then the notes go into a note-taking service, the structure goes into a cloud tool, and the careful part of the arrangement is undone by the boring part of it.

Filamental is the third leg. There is no account to create, no server holding your work and no telemetry of any kind, and your space is Markdown files in a folder you chose. Point LM Studio at it and every element of the arrangement is on the same machine, which is a sentence you can put in front of a client or a security review without qualifying it.

It is also worth saying what this is not. There is no offline mode to switch on, because nothing is online to begin with.

Add Filamental to LM Studio’s mcp.json as an entry named filamental, running node against the bundled server with your space folder passed after a --vault argument:

{
  "mcpServers": {
    "filamental": {
      "command": "node",
      "args": [
        "--no-warnings",
        "/absolute/path/to/filamental-mcp/dist/index.js",
        "--vault",
        "/absolute/path/to/your/vault"
      ]
    }
  }
}

Copy both absolute paths from Settings, then AI Integrations, then Manual setup inside Filamental. You need LM Studio 0.3.17 or later, Node.js 22 or later, and Filamental v0.2.4 or later, on Windows, macOS or Linux.

One prerequisite, and it is the only one that catches people out. The server reads the index Filamental builds and never builds one itself, so the space has to have been opened in the application at least once on that machine. A folder it has not seen stops with a message saying exactly that. If reading a folder somebody sent you is what you want, point the AI at the files directly instead, since a space is plain Markdown and no index is involved.

What the model can then do is exactly what any other client can do: search the space, read notes and properties, follow connections in both directions, create and edit nodes, draw typed connections, and define whole category and connector vocabularies. The connection does not know or care where the model is running.

Building a structure through MCP means calling tools accurately and repeatedly: create this node, give it that type, connect it to the other one in this direction, then do it forty more times without losing the thread. Small models vary enormously at exactly that, and a model that holds a perfectly good conversation about your project can still make a mess of building it.

So try it on a small project first, with something you would not mind rebuilding. If the tool calls come back malformed or the model forgets what it has already created, the honest fix is a more capable model rather than a different configuration, and no amount of prompting will get you round it.

When a local model is not up to authoring, it is often still perfectly good at reading. A structure you built by hand, or with a cloud model, can be interrogated locally afterwards, which keeps the confidential half of the work on the confidential half of the setup.

A short briefing arrives when the connection opens, covering what nodes, categories and connectors are and the conventions that keep a structure readable, and the full skill guide and file format specification are pulled on demand only when a task needs them. That layering matters more with a local model than a frontier one, since a small context window is not something to spend on documentation the job does not need.

Both documents are published on this website word for word as they are sent: the skill file and the file format reference. Nothing else runs underneath them, and the MCP server has no network access of its own.

Free on the Personal plan, which is free permanently with no account and no card, and not credited, seat-limited or reserved for a better tier. The paid tier is $120 a year and buys sending a finished structure to somebody who does not have the application, encrypted on your machine before it reaches us, so we cannot read it either.

If the reason you run models locally is the work rather than the hardware, the In practice pages are written for the fields where that is usually the case.