Onvoi holds transcripts, decisions and internal knowledge, and hands them to software that acts. This page describes the controls around that — and, at the bottom, the ones we do not have yet.
Each collection carries a classification that sets how it is protected and who may reach it. When a record mixes levels, it takes the higher one.
Credentials, OAuth tokens and API keys. Encrypted at rest, never written to logs, never returned in an error.
Transcripts, chat messages, knowledge facts, execution traces and uploaded files. Encrypted in transit, scoped to their owner and organisation.
Task metadata, entity names, skills and org structure. Available to authenticated users within scope.
Marketing pages and published documentation. No protection required.
Authorisation is checked at the route boundary before a handler runs, and again during knowledge retrieval at every hop of the graph.
Short-lived access tokens in HTTP-only cookies, with refresh tokens rotated on use. No server-side session store to steal.
Every scoped route resolves your membership first. Cross-organisation reads are refused at the resolver, not trimmed from a response.
Connected-service credentials are encrypted with AES-256-GCM before they touch the database.
Behavioural inferences about an individual are visible only to that individual — administrators included.
An agent that reads untrusted content and holds real credentials is a genuinely new risk. These are the specific mitigations, not a posture statement.
Incoming task text is screened by pattern matching first and a model only where the result is ambiguous. Repeated attempts against one account escalate and can suspend dispatch.
An agent is given an explicit set of permitted tool classes for a task. A step outside that set is rejected as it happens, not noted afterwards.
Before an instruction goes to an outside agent, sensitive sections are stripped. If the task itself is confidential, it does not leave at all.
Server-side caps on steps, retries, blocked events, runtime and spend apply regardless of how an agent is configured locally.
When a security check cannot run, strict mode blocks the work rather than letting it through unchecked.
Administrators can put every agent in the organisation back to zero autonomy, terminate running work and disconnect them, in one action.
Logs, error reports and internal admin views are the usual route by which customer data leaks sideways. Onvoi masks them at the point of writing rather than trusting reviewers.
These are the third parties that may process customer data on our behalf. Each is reassessed annually and whenever its terms change. Optional services appear only if you switch them on.
| Subprocessor | Purpose | Data processed | Location |
|---|---|---|---|
| Amazon Web Services | Hosting, storage and content delivery | All application data at rest and in transit | United States |
| MongoDB Atlas | Primary database | All application data | United States |
| Anthropic | Language models | Prompts containing meeting, chat and knowledge context | United States |
| Deepgram | Transcription and speaker separation | Meeting audio | United States |
| Recall.ai | Meeting bot recording | Meeting audio, while a bot is in the call | United States |
| Voyage AI | Text embeddings for search | Short text excerpts from your content | United States |
| Stripe | Payment processing | Billing details; no meeting or task content | United States |
Write to security@onvoi.ai with enough detail to reproduce it. We will acknowledge, keep you updated, and credit you if you would like. Please give us a reasonable window to fix the issue before disclosing it publicly, and do not access other people's data while investigating.
Security pages usually list only the wins. Here is what we have not done yet, so it does not come as a surprise in a review.
If any of these blocks an evaluation, tell us — knowing which one stops teams is how it gets prioritised.
We will walk your team through the data flows and answer the awkward questions directly.