Runs on-site
Wide-Learning
Load your firm's past reports, playbooks and notes, even many gigabytes of them. The agent learns from them and answers your team's questions.
For law firms, biotech, healthcare, defence and security teams
Company Agents for confidential work.Your private data stays on your machines.
Ezra-AI runs on your own computers. Local models do the sensitive work on-site. For the hardest steps, private details are filtered out first, and Claude is only used if you approve.
How it works
Most of the work happens on your own computers. Claude is only used for the hardest steps, and only with your approval.
- Step 1
Local models read your data
Ezra-AI runs on your office computers. Local models read your files, scans and logs there.
On-site
- Step 2
Your machines share the work
The agents pool the compute of the computers you connect. That lets them think for longer at low cost.
On-site
- Step 3
Filtered first, then you decide
When a step needs deeper reasoning, a local model removes or abstracts private details. You are then asked if the step may use Claude.
Leaves the building only if you approve
- Step 4
Results come back on-site
Claude's answer returns to your machines. The local agents combine it with your data and finish the work there.
On-site
The path of a request
- Your office computers
- Local filter
- Your approval
- ClaudeOptional
- Back on-site
What the agents do
Each card shows whether it uses Claude.
On-site + Claude, with your approval
Deep-Hypothesis
Give it scan output, logs or a URL. It thinks for several minutes and drafts findings: patterns, anomalies and likely causes.
On-site + Claude, with your approval
Tool-Use
Works with the scanning and analysis tools your team already uses. It learns by watching your team work, then runs routine steps from a written prompt.
Runs on-site
Distributed-Thought
It pools the compute of your office computers at low cost. More compute lets the agent think for longer and give more useful answers.
Runs on-site
Privacy
Private data is processed on your own machines. Your admin decides what, if anything, may be sent to Claude.
Privacy: what stays and what can go
Ezra-AI is built for firms that must keep data on their own premises, or inside the buildings of the clients they work for.
Stays on your machines
By default, your files, scan results, logs, notes and client details are stored and processed on your office computers. Local models do this work.
Can go to Claude, with approval
Only the hardest steps use Claude. A local model first removes or abstracts private details. You are asked to approve the step before anything is sent.
Your admin sets the rules
Your admin decides what may be sent: redacted content, abstracted content, public information only, or nothing. In local-only mode, nothing goes to Claude.
Requests you approve go to Claude through Anthropic's API. To learn how we handle personal data under the GDPR, read our Privacy Policy.
Example use cases
Examples of how a team could use Ezra-AI.
Example · Security
A penetration test at a client's office
- The team runs Ezra-AI on local machines inside the client's building.
- Local models read the scan results and group the issues.
- For a hard finding, the agent asks to use Claude. Client details are filtered out first. The team lead approves or declines.
- The agent drafts the findings report on-site. Client data does not leave the building.
For a stricter engagement, the admin can switch to local-only mode.
Example · Security
An in-house security team
The team runs Ezra-AI on its own office computers. The agents review logs and draft a daily summary. The admin only lets public threat information go to Claude.
Example · Biotech
A research lab analysing trial data
- The lab runs Ezra-AI on its own computers, next to its genomic and clinical data.
- Local models read sequencing results and trial records, and flag patterns worth a closer look.
- For a hard question, the agent asks to use Claude. Sample and patient identifiers are filtered out first. The lead scientist approves or declines.
- The agent drafts a research summary on-site. The raw data stays in the lab.
For the most sensitive projects, the admin can switch to local-only mode.
Example · Healthcare
A private clinic preparing patient summaries
- The clinic runs Ezra-AI on computers inside its own building.
- Local models read patient records and draft visit summaries and referral letters.
- For a complex case, the agent asks to use Claude. Names and other identifying details are filtered out first. The doctor approves or declines.
- A doctor reviews every draft before it is used. Patient records stay in the clinic.
The admin can switch to local-only mode, so no request goes to Claude.
Who it's for
For any team whose data has to stay in the building.
Lawyers
Use AI on client files without them ever leaving your office. EZRA runs on your own servers, so privilege and confidentiality stay fully in your hands.
Biotech
Analyse genomic, clinical and research data with AI that never sends a byte to an outside cloud. Your discoveries and trade secrets stay inside your walls.
Private health providers
Give your doctors and staff AI that works directly with patient records, safely and on site. Patient data stays in your building, under your control.
Defence companies
Bring AI into secure and air-gapped environments where cloud tools can’t go. EZRA runs fully offline, keeping controlled and export-restricted data with your cleared team.
Security firms
Run penetration tests and incident response inside your own building or your client's. Scan results, logs and findings stay on-site, and only filtered requests you approve go to Claude.
Team
Thomas Sichel
CEO
Thomas sets Ezra-AI's direction and leads the work with the teams we build for. He focuses on making private, on-site AI practical for people who handle confidential data.
Tom Erickson
CTO
Tom leads the engineering behind Ezra-AI's Company Agents. He owns the design that keeps sensitive work on local machines and only sends approved requests to Claude.
Andre Luis
AI Infrastructure
Andre builds the infrastructure that runs our local models. He works on pooling compute across office machines, so agents can think for longer at low cost.