The most important setting in a private document search server is the one you never touch: what it can see before you share anything. Most guides to running retrieval on your own computer start with which model to use or how to split documents into chunks. We think the first question is simpler. The moment the server starts, what can a program that holds one of its keys find? The safe answer is nothing, until you say otherwise. We call this the default-closed test, and this page sets out three checks for it, then shows how our own open source add-on, ModelBrain RAG, answers each one.
RAG stands for retrieval-augmented generation. An app looks up the passages that bear on a question and hands them to a language model along with the question. The lookup half can run anywhere. Running it on your own computer means your documents are indexed and searched where they already are, and only the passages your app chooses to send ever reach a model.
That privacy is only as good as the server's reach. A search server that can read your whole home folder, your mail export and your notes is a single key away from handing all of it to whichever script holds that key. So the reach has to be decided by you, one piece at a time.
Three checks, for any local RAG server you are thinking of running:
None of these is exotic. They are the same rules you would want from a shared drive. They are easy to miss with search, because a search server is built to find things, and finding more feels like a feature.
ModelBrain RAG is a free, MIT licensed add-on that gives apps and scripts on your computer an HTTP API for searching documents. The ModelBrain app does the indexing and the ranking. The add-on adds collections, API keys, user accounts and a web page. It contains no ModelBrain code: it starts the helper program the app installed and uses the same tools any AI assistant uses.
| Check | What ModelBrain RAG does |
|---|---|
| Day-one reach | It searches only its own documents folder, one subfolder per collection. Vaults and partitions ModelBrain already holds stay out of reach until an administrator shares them, one at a time, and shared ones are search-only. |
| Limits that travel with the key | Each key has read, write or admin access and can carry a limit list of collections, vaults and partitions. A limited key must name one target in every search and sees nothing else. |
| Refusals that reveal nothing | Asking for something that is not shared, or not on a key's list, gets the same reply as asking for something that does not exist. Results that did not come from the place asked for are dropped. |
Around those three sit the plainer protections. The server listens on 127.0.0.1 only, so other computers cannot reach it, and it makes no outgoing connections. Keys are stored only as SHA-256 hashes and passwords only as scrypt hashes. Requests from web pages on other sites are refused unless you allow that site.
Default-closed only works if sharing is a decision you can see. In ModelBrain RAG an administrator switches each vault or partition on from the web page or the API, and the documentation says plainly what that does: every user and key without a limit list can then search all of it. Mail, chat and saved memories can hold private material, so the documentation asks you to share those with that in mind. We would rather say that out loud than leave it to be discovered.
Before publishing version 0.1 we ran three independent security reviews of the add-on. The first two raised 36 findings and the third, of the installer and first-run setup, raised 15. None was rated high. All were fixed before release apart from three, which are written up in the repository's security notes. The add-on has 69 automated tests, passing on Node.js 18 and 22, and 34 checks run in a real browser against its web page.
On November 14, 2026 we installed it on our own Mac against ModelBrain 0.9.4 with its installer, created the first administrator in the web page, uploaded a document, searched a collection and searched a shared vault.
npm install -g modelbrain-rag, then modelbrain-rag service install and modelbrain-rag open.Then make a key on the API keys tab, and search:
curl -s http://127.0.0.1:8757/v1/search \
-H "Authorization: Bearer mbrag_..." \
-H "Content-Type: application/json" \
-d '{"query": "how long do refunds take", "collection": "handbook", "k": 5}'
The reply lists passages, best first, each with its document, score and headings. The add-on does not write answers and contains no language model. Pass the passages to the model you prefer.
ModelBrain is our product: a memory kept on your own computer that Claude, ChatGPT and Cursor can all use. ModelBrain RAG is a separate, open source piece for a different job. The memory is for what your assistants should know about you and your work. The RAG server is for your own apps that need to search a set of documents, through an API you control.
What is a RAG server? RAG stands for retrieval-augmented generation. An app looks up the passages that bear on a question and hands them to a language model along with the question. A RAG server is the lookup half: you send it a query and it sends back passages, each with the file it came from.
Is ModelBrain RAG free and open source? Yes. The add-on is MIT licensed, with its source at github.com/ModelBrain/modelbrain-rag and an npm package called modelbrain-rag. It needs the ModelBrain app, which is free for local use.
Can other computers reach ModelBrain RAG? No. It listens on 127.0.0.1 only, which is your own computer, and it makes no outgoing connections. Its web page loads nothing from outside your computer.
Does ModelBrain RAG write answers? No. It returns the passages that best match a query, with the file each came from, a score and its headings. Your app passes them to the language model of your choice.
Which computers does it run on? Version 0.1 installs on a Mac with Apple Silicon through a signed, notarized installer, or on macOS and Linux from a command line with Node.js 18 or later.
Last checked November 15, 2026.