Permissions and privacy
Allow, Ask, Deny: the permission model explained
Every operation of every connector carries its own policy (Allow, Ask or Deny), so you decide per action what runs automatically, what waits for approval and what is refused outright; a denial is never worked around.
Last updated 18 September 2026
Three policies, one per operation
| Policy | Behaviour | Good default for |
|---|---|---|
| Allow | Runs without interruption | Read operations you trust |
| Ask | Stops and shows an approval ticket | Anything that writes or sends |
| Deny | Never runs, and no alternative route is attempted | Operations you do not want at all |
How a write is handled
- 1The write must be an explicit request, not an inference.
- 2The target is verified first, the exact channel, repository, branch, page or calendar.
- 3An approval ticket shows what will happen before it happens.
- 4If the result cannot be verified afterwards, it is reported as unverified and not repeated blindly.
Two identical failures stop the loop instead of retrying forever, and cost ceilings stop runaway work.
Credentials and data handling
- Service credentials stay on the backend, bound to your account and encrypted at rest.
- Your AI API key is used server-side and never appears in the conversation, logs or the extension package.
- Provider content is treated as untrusted input, so text inside a page or a message cannot hijack the task.
- Optional website measurement only starts after you accept cookies, and you can withdraw that at any time.
Primary sources
Frequently asked questions
Can the assistant escalate its own permissions?
No. Policies are enforced outside the model, and a denied operation has no alternative route.
What happens if access expires?
The run states that the service is disconnected instead of silently switching to the browser.
Related guides
How connectors work, and when the browser is used insteadConnectors call a service API through the backend; browser actions use the website interface. Learn how resolvenengine decides which path a request takes.Add a custom connector or a remote MCP serverAdd your own HTTPS/OpenAPI service or a remote MCP server as a connector, with owner-bound credential storage, tool discovery and approval policies.Choose a model and keep cost predictableWhy resolvenengine uses your own API key, how to pick a model for a task, and which controls keep token spend predictable.
Try it in your own Chrome
Pro is $19 USD per month and uses your own AI API key, so the provider bills model usage at their own rate.