AI Connections
AI Connection Security
How Ultimate Web Scraper keeps AI agent access safe — workspace-scoped connections, read-only mode, instant revocation, rate limits, and credit protection.
Giving an AI agent access to your data deserves a clear security model. Here's ours.
Two kinds of connections
| Sign-in connections | Access tokens | |
|---|---|---|
| Who uses them | Claude, ChatGPT, and any app that supports OAuth | Cursor and other clients that don't sign in |
| How you approve | On a consent screen — you pick the workspace and the access level | You create and name a token on the Integrations page, picking its access level |
| Where the secret lives | No secret ever passes through your hands | Shown a single time at creation; only a hash is stored |
Both kinds are managed from the dashboard's Integrations page — see Connect an AI agent. A workspace can hold up to 10 connected apps, and a lost token is never a crisis: revoke it and make a new one.
And whichever kind you use, every connection is:
- Workspace-scoped. A connection can only ever reach the workspace it was approved for — never another workspace, and never anyone else's data.
- Instantly revocable. Disconnect an app or revoke a token from the Integrations page and access ends immediately. Connections also stop working automatically if the person who created them leaves the workspace.
- Rate-limited against runaway clients.
Access levels
Every connection is created with one of two levels — picked on the consent screen for sign-in apps, or on the token for everything else:
- Read-only — the agent can analyze tables, query data, and check run status. It cannot change anything or spend anything.
- Full access — adds cleanup operations, starting/stopping runs, and creating new extractions.
Start read-only; upgrade to full access when you trust the workflow. The capabilities page has the full side-by-side table.
Credit protection
- Reading, querying, cleaning, and exporting data consume no credits.
- Starting an extraction returns a cost estimate first so the agent can confirm with you — see AI-started extractions for how that reads in practice.
- Two things charge as soon as you ask for them: mapping a site's URLs (a small flat credit charge, only when the scan actually finds URLs) and re-running an automation you already built (about 1 credit per page, like any cloud run).
- Runs are refused when the balance is empty — balances can't go negative.
Privacy boundaries
- Local data stays local. Agents can only see cloud data. Tables in the Chrome extension's local storage are never exposed — neither is your computer, your files, your open tabs, or your browsing history.
- No cross-workspace access. Team members' other workspaces are invisible to a connection.
- Large-response protection. The server refuses to dump oversized data into a conversation and asks the agent to narrow its query instead.
Security best practices
- Create one token per AI client, named after it ("Cursor — laptop", "Cursor — work machine"), so revocation is surgical.
- Prefer read-only connections for analysis-only workflows.
- Disconnect apps and revoke tokens you no longer use — reconnecting takes seconds.