Capabilities · Permissions
Permissions — approving agent tool-use
When an agent wants to run a tool, you decide. Agentry is local-first: you stay in control of outward and irreversible actions, and the same decision can reach you in three places — the terminal, the Workbench, or over a channel.
A tool call that touches the outside world — or that can't be undone — is gated on your approval. That approval can reach you on three surfaces. They are not three different decisions: they are three views onto the same decision, and the first answer wins.
- The terminal permission prompt — the default. Claude Code asks at the command line and you approve or deny there. This is the baseline path and it is always available; it needs nothing else turned on.
- The Workbench approval surface — when the Workbench is open, a pending tool call shows up as an approval you can act on in the web view, alongside the rest of the run. You stay in the same place where you read and steer the work.
- The channel permission relay — a Channels feature. Claude Code sends a permission request over the channel; you approve or deny it from the Workbench instead of the terminal. The terminal dialog stays open the whole time, and whichever side answers first wins — so the relay never traps a decision in one place.
The posture: local-first, you stay in control
Agentry is local-first. The gate exists so that you stay in control of outward and irreversible actions — the work proceeds on its own where it is safe and reversible, and stops for you where it isn't. The decision is yours by default, not the agent's.
The channel permission relay is a research-preview convenience layered on top — a nicer place to answer the same prompt, never a requirement. Turn channels off and approvals fall back to the terminal dialog, exactly as before; nothing about staying in control depends on the relay being live.
The relay rides the live push layer — see Channels for how that loop works, and Security for the boundary around outward actions.