01
The agent reaches a decision point.
It sends a structured request with the relevant context and response options already attached.
Human decisions for autonomous work
Where agents wait for
your decision.
Agent Outbox is an asynchronous review queue for automated agents. Callers submit questions or approval requests, then retrieve your response when their workflow is ready to continue.
One inbox. Every agent decision that needs your review.
Search · filter · review · respond
01 / Installation
Connect once. Then step away.
Install the CLI and connect an agent to your account. It can send a review request whenever human input is required—even when nobody is watching the run.
- 01
Install with Homebrew
- 02
Connect your agent
- 03
Approve in browser
$ brew install --cask conn-castle/tap/agent-outbox# Connect your agent, then approve it in the browser:$ agent-outbox caller connect my-agentmy-agent is a local label you choose. The connect command opens Agent Outbox so you can review and approve the caller in your browser. Headless sessions use a device code.
02 / The handoff
Agents should ask.
Not guess.
Some steps require human judgment. Agent Outbox makes that handoff a deliberate, reliable part of the workflow, even when the reviewer and agent are working at different times.
02
The review stays organized until you are ready.
Agent Outbox keeps every request visible for the right reviewer, without adding another meeting or live handoff.
03
Your response becomes the next instruction.
The agent receives one explicit response, acknowledges it, and continues from the same point in its workflow.
03 / The decision
See what the agent needs from you.
Each request brings the prompt, supporting material, and available responses into one place. Review the context, respond, and return to your work.
- Agent identity, priority, and revision stay visible
- Approval, choice, text, date, and file responses
- Responses can be undone until agent acknowledgement
04 / Use cases
Put human judgment exactly where it adds value.
Use one review queue for approvals, edits, questions, and operational decisions across your agent workflows.
01
Before it sends
Approve agent-drafted email.
Let an agent prepare the reply and gather the context. Keep final review in human hands before it reaches a customer, partner, or teammate.
02
Before it commits
Confirm a high-impact decision.
Send purchases, policy exceptions, and production changes to the accountable person before action is taken.
03
When a run hits ambiguity
Answer without watching the run.
Let an overnight research or data agent pause on a missing assumption, ask the right person, and resume after the answer arrives.
04
When an automated check fails
Resolve the exception asynchronously.
Route a failed deployment check, unmatched record, or policy exception to an owner, then let the workflow continue from their response.
05 / Wherever you are
Review from any screen.
Review one request between meetings or work through the full queue at your desk. The responsive interface keeps the prompt, context, and actions in view.
Actual responsive product · 430px viewport
06 / Pricing
Start free.
Scale with Pro.
Build and run complete review workflows on the free plan. Choose Pro for higher volume, longer-lived pending requests, file uploads, and expanded storage.
Starter
$0per month
Complete review workflows for everyday volume.
- 5,000 review requests each month
- Up to 1,000 pending requests
- 60-day window for pending requests
- 32 MB review data storage
Pro
$5per month · $50 per year
Higher-volume workflows with files and more storage.
- No monthly submission or queue caps
- Pending requests stay until resolved
- File uploads in human responses
- 1 GB total account storage
Plans apply per account. Both include the complete review queue, search, status controls, and every non-file response type. Standard request-size and rate protections apply.
Ready when you are
Keep agents moving.
Review when it works for you.
Create your free account Free to start · No credit card required