Trust centerEarly access
How Tavix for Work protects company data.
What Tuti can see, what it can do, who approves it, and where your data is stored. Everything on this page describes how Tavix for Work works today. Anything not built yet is marked Planned.
How we protect company data.
The controls below come with every Tavix for Work plan. Admins set them in the admin console, and every change is written to the audit log.
Who can see what
- Two-key access
- Tuti uses a document only when both keys open it: the person can open it in its own app, and its owner shared it to a Space the person belongs to. Both are checked on every search. For confidential documents, access is checked again with the source app right before use.
- Need-to-know hints
- If a document would help but you can’t open it, Tuti tells you who owns it so you can ask for access. It never shows you the title or anything inside it.
- Private channels
- In a private Slack channel, Tuti only uses documents that every member of the channel can open. If it can’t match someone in the channel to a person at your company, it asks to continue in a direct message. In public channels it only uses what’s shared with the whole company.
- SSO and SCIM included
- Sign in with any identity provider that supports OIDC or SAML 2.0, and require it for everyone. SCIM 2.0 adds and removes people and groups automatically. Both come with every plan.
Who does the work
- A separate sandbox for each person
- Each person’s Tuti runs in its own container under gVisor, a sandbox with its own kernel boundary. A firewall blocks it from reaching other people’s Tutis, our internal network and cloud metadata, and it accepts no inbound connections. Inside it, the vault and payment tools only answer Tuti’s own supervised process, checked by the operating system, so code the model runs can’t call them.
- Teammates have their own identity
- AI teammates are accounts of their own, with a name, an owner and their own permissions. They post as themselves, they don’t sign in as a person, and the audit log shows their actions under their own name.
- Scoped, time-boxed delegation
- When a teammate acts for a person, that person grants it a delegation: named tools and resources, for up to 90 days. It can be revoked at any time, and the audit log records each action as “on behalf of” them.
What needs a person
- Approvals based on risk
- Every action gets a risk score from how reversible it is, how much money is involved, who it reaches, how sensitive the data is and how many records it touches. Admins set the thresholds per role, teammate and workflow, and people can only make them stricter. Payments need a passkey or a signed approval on a phone.
- Outside content raises the bar
- Once a run has read something from outside the company, like an outside email, a web page or a guest’s file, its outbound and irreversible actions wait for a person, even in auto mode. Confidential data can’t leave the company in that run without a person’s OK.
- DLP on actions and model prompts
- Built-in detectors find secrets and API keys, card numbers, IBANs, US Social Security and Canadian SIN numbers, email addresses, phone numbers and health terms. For each one, admins choose to allow, redact, ask for approval or block. They check actions before they run, and prompts and replies at our model proxy.
- Browser domain policy
- Admins set the domains Tuti’s browser may and may not open. A teammate’s own list can only narrow the company’s.
- Kill switch
- Stop one teammate, one workflow, one connected app or the whole company. It takes effect within seconds: running work is cancelled and the teammate’s devices are cut off. You can resume it later, and both are logged.
What gets recorded
- Audit log with evidence
- Every tool call, API write and approval is logged, and so is every search: what was read and with whose access. Browser steps keep a screenshot.
- Export to your tools
- Stream the log to an S3-compatible bucket, Splunk, Datadog or your own webhook, signed so you can verify it came from us.
- Retention
- You choose how long the log is kept: 90 days by default, at least 30. Older entries and their screenshots are deleted every day, but never before your export has received them.
- Legal hold
- Turn on legal hold and nothing in the log is deleted until you turn it off. Both are logged.
How data is handled
- No training on your data
- The AI models run on Amazon Bedrock, and Amazon Bedrock does not train models on your inputs. Tavix doesn’t train models on your data either, and our model proxy doesn’t keep prompts in its logs.
- Encryption in transit
- Connections between your browsers, apps and devices and Tavix use TLS, and so do our connections to model providers and the apps you connect.
- Encryption at rest
- Passwords, API keys and app sign-in tokens are kept in a vault encrypted with AES-256-GCM. Each person’s items are sealed with their own data key, and that key is itself encrypted with a master key held in our server’s secret store. Database backups are stored encrypted in Cloudflare R2.
Where your data lives today.
Tavix for Work runs on servers in Germany. The AI models run in the United States. Choosing a region is planned.
Hosting
Hetzner, Falkenstein, Germany
Our servers, databases and each person’s private Tuti.
AI models
Amazon Bedrock, United States
Claude models and text embeddings. Requests go to the us-east-1 region and stay in the United States.
Subprocessors
Companies that process customer data for us.
| Company | What it does | Where | Status |
|---|---|---|---|
| Hetzner Online | Hosting: servers, databases and each person’s Tuti. | Germany | In use |
| Amazon Web Services (Amazon Bedrock) | Runs the AI models and the embeddings for company search. Receives only the parts of a conversation or document a request needs. | United States | In use |
| Cloudflare | DNS for tavix.dev, and R2 storage for encrypted database backups. Our web traffic doesn’t pass through Cloudflare. | Global network | In use |
| Stripe | Billing for paid plans. Card numbers are held by Stripe, never by Tavix. | United States | Pending |
| Resend | Sends email: sign-in codes, invitations, approvals and teammate email. | United States | Pending |
| Expo | Delivers push notifications to the mobile app. | United States | Pending |
Pending means the service is set up but doesn’t process customer data yet.
Apps your company chooses to connect
These only receive data if you connect them. Tavix reads and writes there with the access you grant, and the app processes the data under your own agreement with it.
- Recall.aiPending
Sends a notes bot to meetings you invite it to, and returns a transcript. It names itself as an AI in the meeting.
- SlackIf you connect it
Ask Tuti, teammates and approvals in your Slack workspace.
- MicrosoftIf you connect it
Outlook, Teams and other Microsoft 365 apps you connect.
- GoogleIf you connect it
Gmail, Calendar, Drive and other Google Workspace apps you connect.
Compliance roadmap.
Where we are on certifications and contract terms. Planned items have no date yet.
- Available
Subprocessor list
The list on this page.
- Draft
Data processing agreement
A draft DPA based on GDPR Article 28. Ask us for a copy.
- Draft
AI use and data processing policy
How Tuti uses AI models and your data. A draft is available on request.
- Planned
SOC 2 Type I
An independent audit of how our controls are designed.
- Planned
SOC 2 Type II
An audit of how those controls work over several months.
- Planned
ISO 27001
Certification of our information security management.
- Planned
Independent penetration test
A test of the product by an outside security firm.
- Planned
Your own keys (BYOK)
Encrypt the vault and the search index with keys you control in a key management service.
- Planned
Choice of region
Keep your data in the US, the EU or Canada.
- Planned
HIPAA
A business associate agreement for health data.
Documents for your review.
Send us your questionnaire, or ask for any of these. We’ll send them by email.
- On request
Security review pack
This page in one document, with architecture notes and answers to common questionnaire questions.
- On request
Data processing agreement (DPA)
Draft, based on GDPR Article 28.
- On request
AI use and data processing policy
Draft: which models we use, what they receive, and what they keep.
Opens an email to sales@tavix.dev.
Report a vulnerability.
If you think you’ve found a security problem in Tavix, email security@tavix.dev.
- Tell us what you found and the steps to reproduce it.
- Use only your own accounts and data. Don’t access other people’s data or disrupt the service.
- Give us reasonable time to fix it before you share it publicly.
We read every report and we’ll tell you what we find.
Questions from your security team?
Send us your security questionnaire and we’ll fill it in, or ask for the review pack.