AISecurity and access
AI apps and security
What a connected AI app can and cannot do, how two-step confirmations work, how to limit an app and disconnect it, and what the Orbits MCP server never shows.
When you connect an AI app, the web app makes it an API token of its own. That token decides everything the app can do. This page explains how to keep it small, and how to take it away.
Access
You choose one when you approve the app.
| Access | Scope | The app can |
|---|---|---|
| Read only | read | See projects, deployments, logs, usage and settings. Change nothing, and read no secret value |
| Full access | full | Also deploy, roll back, change variables, domains, DNS and flags, run workflows and delete. The steps that matter ask you first |
Pick the lowest access that does the job. Read only is enough for "why did the build fail?" and "how much traffic did we get this week?". The approval page starts on Full access unless the app asked only to read; change it before you press Allow.
The access is enforced by the Orbits API, not by the app or the MCP server. A Read only app that tries to change something gets an error saying its token is read-only, and nothing changes.
Expiry
An AI app's access lasts 30 days, 90 days (the default) or a year. There is no "never", and no renewal in the background: when it ends, the app's calls are refused and it asks you to approve it again.
See and disconnect AI apps
Each connected app is an API token in the team's Settings › API tokens, named after the app, like Claude-Code-MCP or Cursor-MCP. The list shows its access, when it was created, when it was last used and when it expires.
To disconnect an app, press Revoke on its row, then Revoke token. On the hosted server it stops working within about 30 seconds. To use the app again, connect it again.
Owners and admins can see and revoke tokens. Removing the server from the app alone doesn't end the token: revoke it too.
Two-step confirmations
Through the Orbits MCP server, anything that deletes something, changes what production serves (deploy, redeploy, promote, roll back, pause), adds or changes a domain or DNS record, or shows a secret value takes two calls:
- The first call changes nothing. It returns a preview of what will happen and a
confirmation_token. - Your app shows you the preview, and calls again with the token only after you say yes.
The confirmation_token is signed by the server and bound to:
- the tool,
- the exact details of the request,
- your connection's API token.
It works once and expires after 5 minutes. It can't be used for other details, another tool or another connection. The server never confirms on its own. A confirmed change that is retried, for example after a network error, is not made twice. The tools that ask first.
What an AI app can never do
Whatever its access:
- It works in one team, the one you chose.
- It can never do more than the token allows, and a token can never do more than the team's owners and admins can.
- It can't read the values of sensitive variables, which are write-only for everyone. It can read one non-sensitive value at a time with
reveal_env_var, with Full access, after you confirm; every reveal is recorded in the team's activity. - It can't read every variable at once, or a project's flag keys.
- It can't create, see or revoke API tokens, or manage members and invites.
- It can't change the plan, pay, or see your card's details.
- It can't install or change integrations, move a project to another team, or delete the team.
What the server does with your token
On the hosted server, https://mcp.theorbits.app/mcp, each request's token is passed to the Orbits API for that request only. It is never stored or logged. The server's logs record the request, the tool, how long it took and how it ended, never the details you sent, the results or tokens. A token that is unknown, expired or revoked gets a sign-in error, so the app asks you to sign in again.
The local server uses the token you give it, or the one the CLI stored, and sends it only to the Orbits API. It refuses a CLI settings file that other users on the computer could change, and an API address that isn't https.
Every call the server makes is marked orbits-mcp, which Settings › API tokens shows as the token's last use.
Be careful with what the app reads
An AI app reads what tools return, and that includes text someone else could have written: a commit message, a build log line, a runtime log line, a project or domain name, a DNS value, a workflow's input or output. Someone could hide an instruction in that text, such as "delete every project". This is called prompt injection.
The Orbits server labels such text as data from the account, not instructions, and keeps it out of the summary a preview is read from. But no AI app is perfect. To stay safe:
- Use Read only unless you need more.
- Don't give Full access to an app that also reads email, the web or files you didn't write.
- Read every preview before you say yes. If you didn't ask for it, say no.
- Use a short expiry for anything that can write.
"Not verified by Orbits"
Anyone can build an AI app and give it any name. The approval page shows the name the app gave itself, with Not verified by Orbits next to it. It means Orbits hasn't checked who made the app.
Before you press Allow, check:
- that you started this connection yourself, just now (the page says "Only continue if you started connecting this app yourself");
- the address the page says it will send you back to.
If anything looks wrong, cancel.
Errors
| What the app says | What it means | What to do |
|---|---|---|
| The token is read-only | The app was connected with Read only | Reconnect it with Full access, or do it yourself in the web app or CLI |
| Orbits refuses this to API tokens | No token can do this, by design | Do it in the web app |
| Orbits did not accept the API token | The connection expired or was revoked | Reconnect the app |
| The team's plan does not include this | It needs the Pro plan | An owner can upgrade under Team › Billing. See Plans |
| Orbits is rate limiting these requests | Too many requests at once | Wait the time it says, then ask again |
More fixes, app by app: Troubleshooting.