Trust & privacy
What a connected assistant can do
Before you seat an AI assistant at your desk, here is exactly what it can reach, how the data moves, and how to take the key back.
The short version
When you connect an assistant through the Model Context Protocol, it acts as you, with exactly your permissions, and only within the scopes you approve on the consent screen. Every request it makes runs through the same permission checks as the app itself, enforced on the server — a connection can never see or do more than you can.
Permission matrix
A connection holds one or both of these scopes. You see them named on the sign-in consent screen and approve them before anything is granted.
| Scope | What it allows | What it never allows |
|---|---|---|
| campaigns:read | Read the worlds, sessions, journals, quests, and lore you can already see in the app. | Reading anything you cannot see — GM-only notes and hidden lore stay hidden for a player’s connection. |
| campaigns:write | Create and update journals, quests, lore, and recaps — as ordinary entries under your name, which you can edit or delete like anything you typed. | GM-only actions a player cannot take (publishing a briefing, touching GM notes); deleting your account; billing changes. |
Granting campaigns:write is optional. A read-only connection can answer questions but can write nothing back.
How the data flows
- Your assistant connects to the Storywright MCP server at
https://mcp.storywright.io/mcp. - It sends you to Storywright’s own login and consent screen. You sign in there — the assistant never receives your password.
- You approve the scopes. Storywright issues the assistant a short-lived, scoped access token (PKCE-protected OAuth 2.1).
- When you ask the assistant something, it calls the server with that token. The server re-checks your permissions on every request and returns only what you are allowed to see.
- Anything it writes lands as an ordinary entry under your name, subject to the same reviews and approvals your table already uses.
How to revoke access
- Remove the Storywright connector in your assistant’s settings. Storywright honours standard OAuth token revocation, so a well-behaved client surrenders its token immediately.
- Even without that, tokens expire on their own: access tokens last about an hour and are renewed only while the connection is in use; a connection left idle lapses within thirty days.
- To cut a connection off immediately, email [email protected] and we will revoke it from our side.
Current limitations — as of Storywright, July 2026
We document what is true today, not what we plan. Right now a connection’s scopes apply across all the worlds your account can reach — access is not yet scoped to a single world you pick at connect time. It is still bounded by your own permissions (a player’s connection never exceeds that player’s access), but if you keep several unrelated campaigns under one account, a connection you authorize can reach all of them. Per-world scoped authorization is planned; until it ships, connect from an account whose worlds you are comfortable exposing together, and revoke when you are done. We will update this note the day that changes.
Questions people actually ask
- If a player connects an assistant, can it read my GM secrets?
- No. A connection acts as the person who authorized it, with exactly their permissions. A player’s assistant sees only what that player sees; GM-only notes, hidden lore, and unrevealed entries stay hidden.
- Does the assistant get my password?
- No. You sign in on Storywright’s own screen; the assistant receives a scoped, expiring token instead — and only the scopes you approve.
- Can it delete my account or change my billing?
- No. The connector scopes cover campaign content only. Account deletion, password changes, and billing are out of reach for any connection.
- Do you train models on my campaigns?
- Your campaign content is yours. The connector hands your own assistant a token to read and write on your behalf; it is not a pipeline for us to train on your worlds. See the Privacy Policy for the full data-handling terms.