Skip to content

Approvals

Your agent is powerful, but some actions need a human decision. Approvals let you set up gates so the agent asks permission before doing anything risky — deleting files, running commands, calling external APIs, deploying code.

You approve, reject, edit the request, or ask for changes — all from the dashboard.


requests approval

you approve

you reject

you edit

Agent wants to
delete a file

Dashboard
Approval Panel

Agent proceeds

Agent stops

Agent uses
updated params

  1. Your agent calls a tool that requires approval
  2. The agent pauses and sends a request to the dashboard
  3. You see the request with full details — what tool, what parameters, why
  4. You choose: Approve, Reject, Edit Parameters, or Request Changes
  5. The agent continues (or stops) based on your decision

  1. Open Approval Settings

    Go to the Admin DashboardWorkspaces → select your workspace → Approvals tab.

  2. Create an Approval Config

    Click Add Approval Rule and configure:

    • Tool Pattern — which tools need approval. Use * for all tools, or specific patterns like github__delete_* for all GitHub delete operations
    • Risk Level — classify as low, medium, high, or critical
    • Auto-Approve — optionally auto-approve for specific roles or low-risk patterns
  3. Save

    The rule takes effect immediately. The next time your agent calls a matching tool, it will pause and request approval.


When a request comes in, you see:

  • Tool name — which tool the agent wants to call
  • Parameters — the exact arguments the agent is passing
  • Context — what the agent was doing when it made the request
  • Risk level — low, medium, high, or critical

You have four options:

The agent proceeds with the original parameters. The tool executes and the agent continues its work.

The agent is told the request was denied. It can adjust its approach or ask you what to do instead.

Change the parameters before approving. For example, if the agent wants to delete a directory, you can narrow it to a specific file. The agent uses your edited parameters.

Send a message back to the agent explaining what you want changed. The agent reads your feedback, adjusts its approach, and submits a new request. This can go back and forth as many times as needed.


Approvals are not just yes/no. You can have a conversation about the request:

Agent: “I want to run rm -rf ./old-data/ to clean up disk space.”

You: “Don’t delete everything. Only remove files older than 30 days.”

Agent: “Got it. Updated command: find ./old-data/ -mtime +30 -delete. Requesting approval.”

You: Approve

Each round of discussion is tracked as part of the approval record, so you have a full audit trail of why decisions were made.


Pattern: *
Risk Level: medium
Auto-Approve: none

Every tool call requires your approval. Good for initial setup when you want to understand what your agent does.

Balanced — Read Freely, Write with Approval

Section titled “Balanced — Read Freely, Write with Approval”
Pattern: *__get_*, *__list_*, *__search_*
Auto-Approve: all
Pattern: *
Risk Level: medium
Auto-Approve: none

Read operations (get, list, search) run without interruption. Anything that writes, creates, updates, or deletes requires approval.

Team Setup — Auto-Approve for Experienced Members

Section titled “Team Setup — Auto-Approve for Experienced Members”
Pattern: *
Risk Level: low
Auto-Approve: admin role
Pattern: *__delete_*
Risk Level: critical
Auto-Approve: none

Admins can auto-approve low-risk operations. Delete operations always require explicit approval regardless of role.


If you do not respond to an approval request within 24 hours, it expires. The agent is notified that the request timed out and can either retry or move on.


All approval decisions are logged. Go to WorkspacesApprovals to see:

  • Every request, with status (approved, rejected, expired, changes requested)
  • Who approved or rejected it
  • The full conversation thread for multi-round approvals
  • Timestamps and risk levels

This is your audit trail for compliance and review.