/data-and-ai
Data & AI
How MCP, API keys, and AI clients access SophionOS, who is responsible for what once a client is connected, and how to cut off access.
Connecting an AI client is always your explicit action, with a key you create. We do not call a model provider with your notes.
Last updated: October 20, 2026
Version 2.0 · effective October 19, 2026
Default behavior
SophionOS does not call OpenAI, Anthropic, or any other model host with your notes. There is no background job that ships your workspace to a model. Connecting an AI client is always an explicit action you take, with a key you create.
We do not train models on your workspace content, and we do not use it to improve a model that anyone else runs.
Connection flow
- Create an API key in Settings → AI Access (default: read-only).
- Configure your MCP client (Claude Desktop, Claude Code, Cursor, Codex, Antigravity, OpenCode, or any MCP-compatible client) with the base URL and the key.
- The client calls the SophionOS REST API. SophionOS enforces authentication, access mode, scopes, and your AI-access flags on every request.
- Revoke the key, or use "Disable all AI access", to cut off further requests immediately.
Access modes
- Read-only: query context; cannot create, update, or delete.
- Write-limited: create and update selected entities; no permanent delete or archive-class actions.
- Write-enabled: broader writes including archive; permanent delete remains dashboard-only, so no API key or MCP client can hard-delete your data.
Manage connections in Settings → AI Access (signed in).
Who is responsible for what
Once a client is connected, two things are true at once, and it is worth being precise about the boundary:
- SophionOS is responsible for what it serves and refuses: authenticating the key, enforcing the mode and scopes, isolating your rows, honouring a revoked key, and keeping the activity record.
- You and your client are responsible for what leaves SophionOS. The client and its model provider process whatever they send under their own terms, in their own jurisdictions. We cannot see, instruct, audit, or delete data held by that provider.
That means the privacy policy of the assistant you use matters as much as ours. If an AI provider's terms allow training on inputs, that applies to whatever you let it read.
What an AI client cannot do
- It cannot permanently delete your data, whatever its write mode.
- It cannot keep working after you revoke its key or disable AI access: the next request is rejected.
- It cannot act while global AI access is disabled, even with a valid key.
- It cannot exceed the rate limit of your plan; it will get a 429.
Prompt injection and tool abuse
The API and MCP server are experimental. If you import untrusted content into your workspace, for example a note containing text that looks like instructions, an AI client reading that note may treat it as instructions. This is a property of AI clients, not a bug we can patch from our side. Keep write-enabled keys away from untrusted content, and prefer read-only keys for assistants that browse or import.