The connection runs both ways, and the direction that surprises people is outward. Whatever your Claude has accumulated about a piece of work can leave that window as a structure you can walk around and hand to somebody, rather than as another document. One click in Settings, free on every plan, and the files stay on your own disk.
Prose in a window, or a report, or a deck, or a PDF. That is the entire output range of every assistant any of us uses, and it is why so much good AI work evaporates: the thinking was relational and the artefact is flat, so the next person receives the conclusions without the shape that produced them.
A live connection gives it another option. Ask for the work as a space and Claude writes nodes, notes, categories and typed connections into a folder on your disk, and what you have at the end is navigable rather than paginated. The thing you hand over is the model itself, with the dependencies still in it, instead of a description of a model somebody has to rebuild in their head.
This is also where the weight comes from. We are not the ones who know your field, your client or the last four years of your working life, and your Claude, with its memory and your folders behind it, very nearly is. It has never had anywhere to put that except a message box.
A space is a folder of Markdown files with the relationships written into the front matter, so handing one over hands over something an AI can read natively. The person receiving it points their own Claude at the folder and it is briefed in seconds, with the dependencies, the directions and the reasoning intact, rather than being asked to infer a structure from twenty pages of prose.
That is the file-reading route rather than this connector, and the distinction matters at the receiving end. Reading the folder needs nothing installed at all. Reaching it over MCP needs Filamental, because the server reads an index the application builds.
That closes a loop nothing else in this category closes. Your accumulated knowledge goes out through the connection as a structure, travels as ordinary files, and arrives as input to somebody else’s assistant. Neither end needs an account, a licence or the same software, since a folder of Markdown asks nothing of whoever holds it.
When the recipient is a person rather than a machine, Publisher is the other half of it: the same space as a link or a single HTML file they open in a browser. That one is the paid feature. Handing over the folder is not.
You had a very good session about the project in March. Claude understood the whole thing by the end of it, including the two constraints that make the obvious answer wrong. None of that survived the window closing, so in April you set it out again, slightly differently, and got advice that was subtly worse because one of the constraints didn’t make the retelling.
Pasting the files in doesn’t fix it, because a pile of documents is not the same as knowing what depends on what. The reasoning was never in any one of them, it was in how they sat together, and that part only ever existed in your head and in the conversation you have now lost.
Open Filamental, then Settings from the hamburger icon at the top right, then AI Integrations, then Connect to Claude Desktop. Restart Claude Desktop and it is there.
There is nothing to install and no key to paste. The MCP server ships inside Filamental and is written to disk by the installer, so the button’s whole job is to add an entry to Claude’s own configuration file with the paths already resolved for your machine. If you would rather do it by hand, the exact paths are on the same screen under Manual setup.
What you get after the restart is Filamental listed among Claude’s tools, pointed at whichever space you last had open. Change space later and the connection follows, with no second restart and nothing to reconfigure.
Writing is the direction most people connect it for. Talk a subject through in the way you already do, then ask for it as a structure, and it creates the nodes, writes the notes on them, draws the typed connections between them and can define the categories and connector vocabularies to hold it all. The space builds in front of you in the application, which is also how you catch it misunderstanding something while that is still cheap.
Reading is the half that surprises people. Claude can search the space, open any node and read its notes and properties, and follow connections in both directions, which means it can see that a decision rests on an assumption that rests on a document somebody has since revised. That is the difference between an AI that has your files and one that has your reasoning, and it shows up as a markedly better answer to the same question.
None of that needs a translation layer, since a node is a Markdown file with its relationships written into the front matter, which is a format the model has read a great deal of already.
A briefing arrives the moment you connect, covering what nodes, categories and connectors are and the conventions that keep a structure readable. Beyond that Claude can pull the full skill guide and the complete file format specification on demand, but only when a task actually calls for them, which is a better arrangement than pasting a skill file in by hand because nothing is spent on detail the job doesn’t need.
Both of those documents are published here in full, word for word as they are sent: the skill file and the file format reference. Not a summary and not a friendly description of what they roughly say. If you are going to let software brief an AI on your behalf, you are entitled to read what it says.
Nothing else is running underneath. There is no account, no sign-in and no telemetry anywhere in Filamental, so the only thing that leaves your machine is whatever Claude sends to Anthropic, exactly as it would in any other conversation you have with it.
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.
You need Claude Desktop, Node.js 22 or later, and Filamental v0.2.4 or later. Nothing else, and no separate download for the connector itself.
The one-click button is offered on Windows and macOS, which are the platforms Claude Desktop runs on. Filamental itself runs on Linux as well, and a Linux machine reaches the same server through Claude Code instead, which is hand-configured and works everywhere.
If you would rather not install anything at all, Filamental’s own chat window takes an API key from any of sixteen providers and gives you the same reach without Node, without a second application and without a restart.
The connection is free on the Personal plan, which is free permanently with no account and no card, and it is not credited, seat-limited or reserved for a better tier. We are not in the middle of the call, so there is nothing for us to charge for even if we wanted to.
Every install starts with fifteen days of everything unlocked, and when that ends nothing switches off: the install moves to Personal and stays there. The paid tier is $120 a year, annual only, and it buys the one thing this page has not mentioned, which is sending a finished structure to somebody who does not have Filamental.
Open Settings in Filamental, from the hamburger icon at the top right, go to AI Integrations and click Connect to Claude Desktop, then restart Claude Desktop. Filamental writes the entry into Claude's own configuration file and resolves the paths itself, so there is nothing to install, no key to paste and no JSON to edit. The one-click button is offered on Windows and macOS, which are the platforms Claude Desktop runs on.
Whichever one you last opened in Filamental. The connection follows the open space rather than being pinned to a folder, so switching to a different piece of work needs no reconfiguring and no second restart. Filamental does not strictly have to be running for Claude to reach the files, although in practice you will have it open anyway.
No, because there are no servers in the path. The MCP server is a local process on your own machine talking to the Markdown files in your own folder, and the only thing that leaves is whatever Claude sends to Anthropic, exactly as it would in any other conversation you have with it. Filamental has no account, no sign-in and no telemetry of any kind, so there is no channel for your work to leave by.
Both. It can search your nodes, follow connections in either direction, read the notes and properties on them, and it can create nodes, edit them, draw typed connections and define whole category and connector vocabularies. Talking a subject through and then asking for it as a structure is the ordinary way people use this, and the result appears in the application as it is built.
A short briefing arrives the moment you connect, covering what nodes, categories and connectors are and the conventions that keep a structure readable. Beyond that it can pull the full skill guide and the complete file format specification on demand, but only when a task calls for them. Both of those documents are published on this website in full, word for word as they are sent, so you can read exactly what your AI is told.
Yes, and in two ways. A space is a folder of Markdown files, so you can hand over the folder itself and it opens in any editor, while anyone with an AI can point theirs at that same folder and have the whole structure read natively, relationships and directions included. For a recipient who would rather just look at it, Publisher turns the space into a link or a single HTML file that opens in an ordinary browser with the structure navigable inside it. Handing over the folder is free, and Publisher is the paid feature.
Nothing, on either side of ours. The MCP connection is free on the Personal plan, which is itself free permanently with no account and no card, and it is not metered, credited or seat-limited, because we are not sitting in the middle of it. You pay Anthropic for Claude exactly as you already do, and the only paid feature in Filamental is sending a finished structure to somebody who does not have it.
Connect it, open an empty space, and say so in those words. What comes back is the shape of a thing you already talked through, which is the fastest way to find out whether any of this suits you.