At a glance
Haven is an AI workspace that runs on a computer your firm owns, in your office. Questions, documents, chat history and answers are processed and stored on that hardware. Haven does not use ChatGPT, Claude, Gemini or any other outside AI service, and it doesn't need an account with one.
On your hardware
The AI model runs on a computer in your office. Your firm owns it.
No outside AI
Cloud AI models are blocked in the app. Prompts and files are never sent to an AI provider.
Encrypted at rest
Each person's chats and files are encrypted with AES-256, unlocked only by their password.
Two-step sign-in
Codes from an authenticator app. Your firm can require it for everyone.
Sealed audit log
Who signed in, what they asked and which files were used. Only your admin can open it.
Answers cite sources
Answers about your documents point to the exact passages they used.
Where your data goes
Everything a person types or uploads stays inside your office network. The only internet traffic Haven needs is downloading software and model updates, which come in. Nothing about your work goes out.
| Data | Where it lives | Leaves your network? |
|---|---|---|
| Questions (prompts) | Your Haven hardware | Never |
| Uploaded documents and the text pulled from them | Your Haven hardware, encrypted | Never |
| Chat history and answers | Your Haven hardware, encrypted | Never |
| Audit log | Your Haven hardware, sealed to your admin | Never |
| Software and AI model updates | Downloaded from the publishers | Download only |
| Your contact and billing details | Our business records | Covered by our privacy policy |
Network posture
- Local model only. Haven talks to an AI model running on the same computer or on a server inside your office network. It refuses to connect to a model server outside your network.
- Cloud models blocked. Some local AI tools can quietly route requests to cloud models. Haven hides and refuses them.
- No web requests from the app. The Haven window can't load anything from the internet. Any attempt is stopped and written to the audit log.
- Visible status. A privacy indicator in the app shows where the model is running, that cloud models are blocked, and how many outside requests were stopped.
- Remote work over a private VPN. Staff outside the office connect to Haven through a private VPN we set up. Haven isn't exposed to the open internet.
- Works offline. Haven keeps working if your internet connection goes down.
Encryption
- Per-person encryption. Each person's chats and documents are encrypted with AES-256-GCM. The key comes from their password (scrypt), so files on disk can't be read without it.
- No back door. There's no master key and no password reset that bypasses encryption. Not even we can open a person's chats. The trade-off: a forgotten password means that person's history can't be recovered.
- Full-disk encryption. We turn on macOS FileVault during installation, so the whole drive is protected if hardware is lost or stolen.
- Two-step secrets protected. Each person's authenticator secret is encrypted with their own password key.
Sign-in and access
- Individual accounts. Everyone signs in as themselves. No shared logins.
- Two-step sign-in. A 6-digit code from Microsoft Authenticator, Google Authenticator, 1Password, Authy or any standard app. Codes can't be reused, and repeated wrong codes end the attempt.
- Recovery codes. Eight single-use codes for a lost phone, shown once and stored only as hashes.
- Firm-wide requirement. Your admin can require two-step sign-in for everyone. People without it set up are asked to enroll at their next sign-in.
- Lock anytime. One shortcut locks Haven and clears the session key from memory.
Audit log
Haven keeps a record that helps your firm supervise how AI is used.
- What's recorded: sign-ins and failed attempts, each question asked, the model used, the files in the conversation, which passages answers cited, files added or removed, deleted chats, security setting changes and admin actions.
- Sealed: every entry is encrypted to your firm's audit key. Staff can't read the log, including their own entries. Only someone with the admin passphrase can open it.
- Tamper-evident: entries are chained together, so edits, deletions or reordering inside the log are flagged when it's opened.
- Exportable: your admin can export the log to a spreadsheet for review, a client or an insurer.
The admin passphrase is set by your firm at installation. We don't keep a copy.
Checking answers
AI can be wrong. When Haven answers a question about your documents, it numbers the passages it relied on and shows them under the answer, so the person can check the source before relying on it. Your professional judgment still decides what goes to a client.
Updates and support
- Updates you can see. We test app and model updates before installing them, and we tell you what changed.
- Support without seeing your work. Our support process is designed so we never need to open your chats, documents or audit log. Remote sessions happen only when you start or approve them, on screens you control.
- Local help. We're based in San Diego and can come on site.
Incident response
If we learn of a security issue that affects your installation, we'll tell you within 72 hours, explain what we know, and work with you on next steps. Your firm decides whether and how to notify clients or regulators, and we'll provide the information you need to do that.
How Haven supports your obligations
No software makes a firm compliant on its own. Haven is designed to support the obligations below. Your firm stays responsible for its policies, training and professional judgment.
| Rule or guidance | What it asks | How Haven helps |
|---|---|---|
| ABA Formal Opinion 512 (Model Rules 1.1, 1.6, 5.1, 5.3) | Understand the tool, protect client confidences, supervise lawyers and staff using AI | Client files stay on firm hardware with no outside AI provider; audit log supports supervision; cited answers support verification |
| California Rule 1.6 and Bus. & Prof. Code §6068(e) | Maintain client confidences | No third-party AI service receives client information |
| FTC Safeguards Rule (16 CFR Part 314), for tax preparers and other financial institutions | Access controls, encryption, multi-factor authentication, activity logging, service-provider oversight | Per-person accounts, AES-256 encryption at rest, two-step sign-in, sealed activity log; Haven never receives your customer data |
| HIPAA Security Rule (45 CFR 164.312) | Unique user IDs, audit controls, encryption, person authentication | Individual sign-in, audit log, encryption at rest, two-step sign-in; PHI stays on your hardware |
| SEC Regulation S-P | Safeguard customer records and oversee service providers | Customer information isn't shared with an AI vendor |
Every installation comes with a Compliance Kit: an AI use policy, a written information security program (WISP) addendum, client disclosure language and a one-page data-flow sheet you can hand to clients and insurers. They're templates for your counsel to review, not legal advice.
Shared responsibility
| We handle | Your firm handles |
|---|---|
| Installing Haven, FileVault and the VPN | Strong, unique passwords for every person |
| Blocking cloud models and outside connections | Turning on, or requiring, two-step sign-in |
| Testing and applying updates | Keeping the admin passphrase safe |
| Support that doesn't touch client content | Physical security of the office and devices |
| Telling you about security issues quickly | Reviewing AI output before relying on it, and your AI use policy |
Report a security issue
Email andreaus@tryhaven.ai. We'll confirm we received it within two business days. For questions about this page or a security questionnaire, book a call and we'll walk through it with your IT provider.