A confidential assistant for lawyers, accountants and anyone else bound by client confidentiality. The model stores nothing and learns nothing from you.
One prompt, start to finish
Every place your data touches, how long it stays there, and who can see it. The privacy policy repeats this table with the full detail.
| What | Where | How long | Who can see it |
|---|---|---|---|
| Prompts and answers in transit | Between your browser and our server, through Cloudflare | Only while sending | Cloudflare, which decrypts traffic briefly to route it and block attacks. It does not log what you send. |
| Model processing | ZDR serverless inference providerA combination of H200s and serverless inference providers. | In memory, during the request | Nobody. Not stored, not used for training. |
| Chat history and uploaded files | Our server, Germany | Until you delete them | You. There is no admin chat viewer. |
| Server logs | Our server, Germany | 14 days | Us, for security. They contain no chat content. |
| Backups | Encrypted storage, Germany | 30 days, then gone | Nobody without the backup key, which is kept off the server. |
Delete a chat and it leaves the server at once, and the backups within 30 days. Files you upload stay in your file list until you delete them there.
It does the first pass. You keep the judgment.
“Summarize this 40-page shareholders' agreement and flag anything unusual in the transfer restrictions.”
You get a one-page summary with clause references, and a short list of provisions worth a second look.
“Turn these audit meeting notes into a draft management letter, using our usual structure.”
You get a first draft in your format, with open questions marked for you to resolve.
“Rewrite this advice for a client who isn't a specialist, in under 300 words.”
You get a plain-language version to edit and send.
It runs a leading open-weight model. On long documents and professional work tasks it matches or beats two leading closed models from earlier this year. On the hardest reasoning questions, they still lead.
| Benchmark | Our modelOpen weights, maximum reasoning effort | Claude Opus 4.7Anthropic (released 16 April 2026) | GPT-5.5 (xhigh)OpenAI |
|---|---|---|---|
| Long-document reasoningAA-LCR v1.1 | 84% |
79% |
84% |
| Professional work tasksGDPval-AA v2.1, Elo rating | 1600 | 1338 | 1336 |
| Hard reasoningHumanity's Last Exam | 39% |
42% |
46% |
Higher is better. Scores from Artificial Analysis, retrieved 28 September 2026. Our model scored at maximum reasoning effort.
These certifications belong to our providers. Our own controls are listed alongside.
ABA Formal Opinion 512 expects lawyers to understand how an AI tool handles client data before using it, and accountants carry similar confidentiality duties. Every pilot member gets a one-page data-flow sheet to share with clients or counsel.
No tool settles your obligations on its own. The sheet gives you the facts to decide.
Not through the app. There is no admin chat viewer, and every admin sign-in to the server is logged. We run the server and hold its keys, so we won't claim it's technically impossible.
Stored on our server in Germany. The model runs on zero-data-retention (ZDR) inference on H200s in the USA. Traffic passes through Cloudflare, web searches go to Tavily and public search engines, and sign-in emails are sent by Mailgun. Nowhere else.
No tool does that by itself. It gives you verifiable facts about where your data goes, which is what your own assessment needs.
A leading open-weight model, released under the MIT license, running on ZDR inference on H200s in the USA. We'll tell you before we ever change it.
No. DeepSeek publishes the model's weights, and we run them on our own inference setup. DeepSeek the company never receives your prompts.
You can export your history or have it deleted. Deleted data leaves the backups within 30 days.
Seats go to people in our network first. Tell us who referred you.
We use these details only to reply to you.