Control what your AI teammates can do

Decide which actions your AI teammates can take on their own, how trust carries forward, and how to review what they've done.

Helio lets you control what your AI teammates do on their own and what waits for your sign-off. It isn't a single on/off switch — trust is decided per action, as the teammate works.

Live preview

How control works

Your teammate gets on with routine work — reading, searching, reasoning, drafting — without asking. When an action reaches outside Helio or can't be undone, they pause and request approval instead of proceeding.

When they pause, you either:

  1. Approve — the teammate takes the action
  2. Deny — the teammate stops and is told why

Lower-risk work runs on its own; higher-impact actions pause for you:

  • Read actions — reading files, reading emails, searching the web → usually run on their own
  • Write actions — editing files, sending internal messages, creating tasks
  • Sensitive actions — sending external email, running code, calling external APIs
  • Irreversible actions — deleting, publishing, deploying → pause for approval

Where approvals appear

When your teammate needs approval, the request shows up in two places — an approval card in the channel they're working in, and your Inbox. Each request shows:

  • Who is asking and what action they want to take (for example, "Send email to alex@example.com")
  • Why — what they're trying to accomplish

You respond on the approval card in the teammate's channel; opening the request from the Inbox takes you to that card, and your decision syncs back everywhere. Choose Approve or Deny — denying tells the teammate you declined, so they can ask what to do instead.

How trust carries forward

Routine approvals are one-at-a-time by default: approving an action now doesn't grant blanket permission for that kind of action later. Standing, longer-lived trust applies in two narrow cases:

Requests that carry a remember window

Some approval requests come with a short remember window, and the card says so — for example, that the action will be remembered for a few minutes. Approve one of those and Helio skips re-asking for that exact action until the window runs out. The duration comes with the request; it isn't something you pick on the card.

Requests to use a credential

If you publish a credential as requestable in the Vault, a teammate can request it and you Approve or Deny the request. The access policy — trusted (every fetch is allowed) or one-time (released once, then consumed) — comes with the request, not as a choice on the approval. (When you share a credential proactively, you pick the policy yourself.) You manage and revoke these grants from the credential's detail view in the Vault.

Reviewing history

Every reply in a teammate's channel or DM carries its work record inline — expand a reply's steps (Show steps on the message) to replay what they did on that turn: thinking, files read, edits made, commands run, with each step's status.

  • Task activity log — each task's detail panel has its own log of everything that happened on that piece of work
  • Inbox — shows a history of all approval requests (approved, denied, timed out); filter by teammate name

Each log entry includes a timestamp, the action type (e.g. "Sent email," "Edited file," "Ran command"), the specifics, and how it was authorized — autonomous, approved once, or always trusted.

If a teammate did something you don't want repeated, remember that routine actions are approved one at a time — there's no standing permission to undo. The one exception is access to a Vault credential you've shared, which you revoke from the credential's detail view in the Vault.

Ready to try this with your own AI teammate?
Start free with 1,000 credits — no credit card required.
Try Helio →

On this page