Private by default. Clear by design.
Security isn't only encryption. It's the feeling that unfinished thinking, private reasoning, and project context aren't being casually exposed. This is how we build for that.
No token bypasses review.
AI-generated, imported, and agent-proposed memory enters review before it becomes trusted. Human-approved project edits are audit logged. Retrieval surfaces only read the trusted set.
Nine commitments. Kept, not advertised.
Workspaces are yours
Every workspace is owned by the person or team who created it. Imports, prompts, sources, decisions, and notes stay scoped to that workspace and never bleed into anyone else's.
Sharing is opt-in
Nothing is shared by default. When you do share (a comparison, a project export), the share is scoped, revocable, and visible in the audit log.
Clear ownership
Your work is yours. Export at any time. Delete at any time. Shelvia doesn't hold your data hostage and doesn't use your imports to train models.
Sensible access controls
Workspaces support roles: owner, admin, member, viewer. Roles control who can edit, share, or invite, so a teammate who needs read access doesn't accidentally archive a project.
Imports are treated as untrusted data
Pasted content from chats, screenshots, and links is treated as data, not as instructions. Shelvia doesn't obey commands hidden inside imported text, a real risk in any tool that handles AI output.
Encryption in transit and at rest
Connections to Shelvia use TLS. Files and data at rest are stored encrypted. We don't store credentials in client-side code.
Archive before delete
Most destructive actions go through archive first. Recoverable on demand. Permanent deletion requires explicit confirmation.
Exports without lock-in
Every project supports Markdown and HTML export. The point isn't the format. It's that your project memory leaves with you.
Multi-provider AI, with clear scope
Shelvia routes AI calls to a small set of named providers: Anthropic and OpenAI by default, with Perplexity used only for tasks that benefit from live web search. Each call is scoped to one task, providers see only the content that task needs, and provider choice is logged so you can review which model handled what.
What we don't do
A few specific things we've seen in this category that Shelvia intentionally avoids.
- We don't train models on your content.
- We don't make workspaces public by default. Every share is explicit.
- We don't claim certifications we don't have. When we have them, we'll show them honestly.
- We don't use phrases like “military-grade encryption.” Marketing language doesn't make TLS more secure.
- We don't sell your data, share it with third parties for advertising, or use it to build a profile of you across workspaces.
- We don't hold your data hostage. Exports are part of the product, not a paid upgrade.
Security FAQ
Specific questions, specific answers. If something here is missing, email team@shelvia.net.
Does Shelvia train on my content?
No. Your imports, prompts, sources, decisions, and notes stay inside your workspace. They are not used to train models, ours or anyone else's. Workspace adaptation (better source-detection over time) is private to your workspace and never shared across tenants.
Who can see my workspace?
Only members you've invited. The default state of every workspace is private. You can promote a member to admin to share invite-management, but adding members is always an explicit action, never automatic from email domain unless you've set up SSO with domain-based auto-join.
What happens if I delete my account?
Your workspaces, imports, prompts, sources, decisions, and notes are deleted. We don't retain a copy "just in case." If you're part of a shared workspace as a non-owner, your account leaves the workspace; the workspace itself stays for the other members. We recommend exporting first if you want to keep anything.
How does sharing a comparison work?
When you share a comparison, Shelvia creates a share link with a unique token. The link is scoped to that one comparison; it doesn't grant access to anything else in your workspace. You can revoke a share link at any time, and revocation is immediate. The audit log records who shared what and when.
Are my imports treated as instructions to the AI?
No. This is a real risk for any AI tool that processes user content. A malicious chat or webpage can include hidden instructions intended to manipulate the model. Shelvia wraps imported content in explicit untrusted-data tags before any AI processing. The model treats imports as data to summarize, not commands to obey.
Where is my data stored?
Shelvia is hosted on commercial cloud infrastructure with encryption at rest. Specific region details are available on request for paid plans where data residency matters.
Do you have certifications like SOC 2 or ISO 27001?
We will be transparent about certifications when we have them. Today, Shelvia is a small product. We have privacy and access controls in place, but we don't carry formal third-party certifications yet. If you need them for procurement, contact us; we'll tell you where we are honestly.
Can I see who did what in my team workspace?
Yes. Important actions (adding a member, changing a role, sharing a comparison, archiving a project, changing the plan) are recorded in the workspace audit log. Admins can review the audit log to reconstruct activity.
What about "sensitive" imports?
When you mark an import as sensitive, Shelvia skips AI processing on that item. It stays inside your workspace as raw content; the agent and learning systems don't see it. Useful for confidential interviews, legal text, or anything that shouldn't round-trip through a model.
Which AI providers does Shelvia send my data to?
Shelvia uses Anthropic (Claude) as the default provider for summarization, extraction, comparison, and project memory. OpenAI is the fallback when a Claude call fails transiently, and is also the embeddings provider for semantic search. Perplexity is used only for tasks where live web search adds value (for example, source verification or research refresh), and only when a workspace has opted in by configuring the key. Every AI call records which provider handled it, so you can review provider mix in your workspace activity.
Trust shouldn't require legal language.
If you're evaluating Shelvia for a team or a regulated workflow, we'll talk to you specifically about your constraints. No template responses.