Private, local AI for confidential documents — runs on your own machine

The problem isn’t the AI. It’s where your files go

You signed the NDA. Your name is on the engagement letter, the data-processing clause, the promise you made to a client or a co-founder that their material stays private. When something goes wrong with that material, the person who answers for it is you — not a vendor’s status page.

So the offer to “just paste it into a chatbot” is not a convenience. It is a decision you have to be able to defend. Uploading a confidential file to a hosted AI service sends that file to a third party’s servers, where its handling is governed by their terms, their retention policy, and their access controls — not yours. Even where a provider says inputs are not used for training, you are still trusting a claim you cannot inspect, about infrastructure you cannot see, on data you are personally answerable for.

SealedBrief takes the other path. It is a desktop AI workspace that indexes your documents and answers questions across them, and it runs the model on your own computer. Your documents never leave the machine. That is the whole design goal: local AI for confidential documents, where “local” is a property you can check rather than a promise you have to take on faith.

What “local-first, air-gapped” actually means here

Most tools that say “private” mean the connection is encrypted on the way to their cloud. This is different. SealedBrief keeps the documents on your disk and does the thinking there too.

Internally it runs as two operating-system processes that cannot swap roles — a split-brain design that makes the privacy boundary structural rather than a setting:

• The Compute Plane is the part that touches your files. It does the model inference, the retrieval, the search index, and the on-device text extraction. It has no network egress by construction — it does not open outbound connections at all.
• The Presentation Plane is the window you interact with. It reaches sealedbrief.com for two narrow purposes: validating your licence and checking for updates. Neither of those requests contains your documents or your questions.

This is the honest version of the claim, and it is the one worth repeating: it is not that the application makes no network calls ever. It is that the process handling your documents makes none, and the process that does talk to the network never has your content to send. That distinction is what lets SealedBrief run genuinely air-gapped, on-device AI over your own files — the Compute Plane has no network egress — while still shipping licence checks and updates like normal software.

How it answers questions across your own files

Point SealedBrief at a folder. Its ingestion engine indexes the documents in place — PDF, DOCX, TXT, Markdown, EPUB, HTML, and scanned images through on-device OCR — without copying them into a proprietary cloud store. Then you ask questions in plain language, and it retrieves the relevant passages from your corpus and answers from them.

Answers are grounded in retrieved passages from your own documents and show the source paragraphs alongside the response. This is offline, self-hosted document AI with citations: you can see exactly which lines the answer rests on. And when the model cannot ground a claim in your corpus, it says so rather than inventing a citation to fill the gap.

Be clear-eyed about what the model is. The answer is a draft, not a verdict. Treat every response the way you would treat a junior associate’s memo — read the cited passages before you rely on it. That habit is the point of showing the sources in the first place.

How to verify nothing leaves your machine

You do not have to trust the previous three sections. You can watch the traffic yourself.

Start a packet capture on your machine, open SealedBrief, index some documents, and run a full question-and-answer session. Watch what the Compute Plane sends while it works on your files: nothing goes out. You will see just one kind of network activity from the application — the Presentation Plane’s licence and update checks — and you can inspect those requests to confirm they carry no document content or query text.

This is a five-minute check, and it is reproducible. If you have an IT or security person, hand it to them; the result is the same on their machine as on yours. That is the difference between “we don’t send your data anywhere” as a marketing line and as something you can falsify. Here, you can try to falsify it. This is how SealedBrief lets you run AI without sending documents to the cloud and then prove it, instead of asking you to believe it. For the exact commands — the packet capture, the per-process socket checks, and the at-rest entropy test — follow the full verify-it-yourself walkthrough.

What’s encrypted, and what honestly isn’t

At rest, SealedBrief encrypts the sensitive parts of its own store: the extracted text, source references, and metadata are protected with field-level AES-256-GCM, and the core database and full-text search index sit behind SQLCipher’s AES-256 page encryption. Keys are held in your operating system’s keychain.

There is one carve-out worth stating plainly, because leaving it out would be dishonest. The embedding vectors — the numeric representations used to find relevant passages — are stored unencrypted. Similarity search needs to do plaintext math over those numbers, so they cannot be encrypted at rest without breaking the retrieval that makes the product work. Text and metadata are encrypted; the vectors are not. If a page tells you everything is encrypted, it is glossing over exactly this.

The honest limits

A tool you are trusting with confidential work should tell you what it can’t do:

• The model is local, not frontier-scale. SealedBrief runs Qwen3 (an Apache-2.0 open model) on your hardware — not a large cloud-hosted frontier model. Its output is a cited draft you verify, and it will not match a big hosted model on every task. That trade is the price of keeping your documents on your machine.
• It needs real hardware. It runs well on a 12 GB-or-larger NVIDIA GPU, and it also runs on Apple Silicon Macs (M1–M4) via Metal. CPU-only works but is meaningfully slower — tens of seconds per answer. Plan on about 16 GB of RAM and roughly 15 GB of free disk.
• It’s a ~15 GB download. The models ship inside the app, so it is a large one-time download and then it works fully offline — nothing is fetched at runtime.
• No audio or video transcription. It reads documents and images (with on-device OCR), not recordings.
• Platforms: Linux (a portable AppImage that runs on any distribution — no install step) and macOS (a signed, notarized build for Apple Silicon only) ship today. Windows is a waitlist — it is genuinely blocked on an encryption-library dependency, and you are not charged until it ships. If you are on Windows, you can add your name and wait rather than being sold something that does not run yet.

Choose your field

The general case is above; some duties have their own shape. If your confidentiality obligation is specific, start with the page written for it:

• For journalists protecting sources and notes
• For lawyers with privileged client files
• For clinicians and patient records
• For therapists and confidential session notes
• For researchers under NDA
• For consultants and advisors under client NDAs

Each covers the same architecture through the lens of one professional duty.

What it costs

SealedBrief is a 3-year term licence — not a perpetual licence, and not a subscription:

• Personal — $199 for the 3-year term, plus $49/yr support in years 2 and 3 ($297 total across the term). One device.
• Professional — $499 for the 3-year term, plus $109.99/yr support in years 2 and 3 ($718.98 total across the term). Up to three devices, a commercial-use grant that covers paid client work, and an exportable compliance-and-audit evidence surface (at-rest encryption proof, network-egress attestation, and an ingestion log).

Year 1 includes all updates and support in the purchase price. The licence is an Ed25519-signed file verified offline, so activation works even if the company were to disappear — it keeps working for the rest of your term. If a term lapses, the vault opens read-only; nothing is deleted. Every purchase is backed by a 30-day, no-questions refund, which is separate from the free trial below. You can compare both tiers on the pricing page.

Try it on your own documents

The most direct way to judge an AI that keeps your data on your machine is to run it against your own files and watch the traffic while you do.

The download is the full 14-day trial, not a limited demo. On first launch you choose “Start 14-day trial” on the welcome screen — no account, no credit card, and starting the trial makes no network call. For 14 days you get the complete workspace over your own documents: index, ask, read cited answers, search, and export.

After 14 days the app opens read-only. Nothing is deleted and nothing is locked away — you keep full read, search, and export (Markdown and PDF) of everything already indexed. Only indexing new documents and asking new questions pause until you enter a licence, which unlocks the same install instantly. No reinstall, no lost work.

Download the free 14-day trial — the full workspace over your own documents, no credit card. Linux and macOS (Apple Silicon) · ~15 GB download.