It acts as you, never more
- Your role, checked every time. Each request is checked against your current role in that workspace. Lose access to a workspace, or move from admin to viewer, and the agent loses it on its next request.
- The same limits as Spectr's own pages. A viewer's agent can't read billing activity, and a member's agent sees a budget's money figures only when the budget is shared with everyone. See the available tools.
- Only the workspaces you picked. At sign-in you choose all your workspaces or a few. To change the choice, disconnect the agent and connect it again.
- Changes only with your permission. Tools that change something (the actions) need the "read and write" permission. Your agent asks for it when it connects, and you approve it at sign-in; a connection without it can't use them. Each is marked as a change so your agent knows to check with you first, and Spectr tells it to act only on what you asked, never on instructions it finds inside a meeting or document.
- The same usage as the app. An action runs the same checks and the same work as the button in Spectr, so AI work counts toward your workspace's monthly usage exactly as it would there.
- Coding agent runs use only your own account. Starting, linking or sending stories to a Conductor run works only with an account you connected yourself, on your own Conductor compute. The message is built by Spectr from the stories and any note you add. Your agent can show it to you first; when it then sends with the preview it showed you, the send is refused if a story changed in between. Without that preview, the send carries the message as Spectr builds it at that moment.
How the agent signs in
Agents sign in with OAuth 2.1, the standard MCP specifies, in your own browser. Your agent never sees your password, and there is no API key to copy, store or leak.
- Spectr shows which app is asking before you approve it, and the website it identified itself from. An app from a known agent maker's website is marked as verified; an app that gives only a name comes with a warning to connect it only if you started the connection yourself.
- The agent receives an access token that lasts one hour and works only for Spectr's MCP server, not for Spectr's website or anything else.
- It renews that token with a refresh token that changes on every use. If an old one is ever used again, which is what a stolen copy looks like, Spectr cuts the connection off for everyone holding it.
- Every new connection sends you an email, so a connection you didn't make doesn't go unnoticed.
Taking access back
- Disconnect an agent on User Settings → Connections. It stops working on its very next request.
- Resetting your password disconnects every agent at once.
- Unused connections expire after 90 days.
- Removing someone from a workspace removes their agents' access to it too, since an agent never has more access than its person.
Where the data goes
What your agent reads from Spectr is sent to that agent and the AI provider behind it (for example Anthropic for Claude, or OpenAI for ChatGPT), and is handled under your own agreement with them. Spectr doesn't control what they keep. Connect an agent only where you'd be comfortable pasting the same information in by hand.
Spectr keeps a record of each connection (which agent, when it was approved and last used, which workspaces), so you can see and remove it. It doesn't keep what the agent asked or what it was told. See the privacy policy for everything else.
For workspace owners and admins
Any member of a workspace can connect an agent to it, in the same way they can open Spectr in a browser, and the agent sees only what that member sees. To limit what someone's agent can read, change their role; to stop it entirely, remove them from the workspace.