Skip to content
teamctl

Human-in-the-loop (HITL)

teamctl agents can act autonomously on low-risk work. Anything brand-sensitive pauses and asks you first — over any configured interface (Telegram, Discord, iMessage, CLI, webhook).

Three layers of safety, outside-in

  1. Per-agent autonomy — full, low_risk_only (default), or proposal_only.
  2. Globally sensitive actions — a list in global team-compose.yaml. Any tool call tagged with one of these is routed through request_approval.
  3. Auto-approve windows — pre-authorize specific actions in a scope for a bounded time.

Default sensitive actions

publish, release, payment, external_email, external_api_post, merge_to_main, dns_change, deploy.

Configure in global compose:

hitl:
  globally_sensitive_actions:
    - publish
    - release
    - deploy
  auto_approve_windows:
    - action: publish
      project: blog
      scope: "social-media-launch-v3"
      until: "2026-04-25T18:00:00Z"

Agent flow

agent    → request_approval(action, summary, payload?, ttl_seconds?)
team-mcp → insert row in `approvals` (status: pending), block caller on long-poll
interface adapter → surface to you on Telegram/Discord/iMessage/CLI
you      → Approve / Deny / Why?
team-mcp → update status, caller unblocks, returns { id, status, note }

The default is deny until explicitly approved. Requests expire after the agent-supplied ttl_seconds (30s–3600s, default 900s).

Deciding

From any of:

  • Inline buttons on the interface that notified you (Telegram/Discord/iMessage).
  • teamctl pending / teamctl approve <id> / teamctl deny <id> on the host.
  • The approvals table directly (scripts, audit tools).