Human approval
Updated on 3 November 2026.
Require user approval for tools that modify data or systems. With confirmation on, the gateway pauses each call to the tool and asks the user to approve it. The call reaches the MCP server only after the user clicks Approve.
Who approves
The approval comes from the person who made the call, in their own AI app. It is a confirmation step that slows down risky actions and records a justification. It is not an independent approval:
- No administrator, manager or second person reviews the call, and there is no approval queue or delegation.
- A message such as "An administrator requires confirmation for this tool" means an administrator turned the rule on, not that an administrator approves each call.
- By default, anyone who has the Approve in Quilr link can decide, and the decision is recorded as the user who made the call. Treat the link like the conversation it came from and don't share it. To require the approver to be signed in as that user, see Signing in to approve under Limits and good to know.
This feature doesn't provide two-person approval. To reduce exposure for sensitive tools, restrict them instead: disable it with Tool visibility, or limit who can reach it with Group and user rules and Server access.
Choose the instructions for the mode that governs your MCP Gateway. See Switching from classic settings for what happens to stored settings when the engine is on.
- Classic settings
- Policy Engine
Configure it on the server
Each tool has two switches.
- Turning Require confirmation on also turns Require justification on. Turn justification off per tool if approval without a written reason is sufficient.
- Require justification needs confirmation on first.
- The confirmation switches are locked while a tool is disabled.
Click Save settings to apply. Good candidates are tools in the Write tools and Destructive tools groups, such as sending messages, deleting records or merging code.
Different rules for groups and users
Confirmation can differ by smart group or user. In Group and user rules, click Add rule, choose A smart group or A single user, and under Tool overrides set Confirmation and Justification for each tool.
Setting Confirmation to On also sets Justification to On. If a user is in several groups with rules, a group that requires confirmation takes precedence; a rule for that single user applies last. Use Effective settings preview to check one user's effective settings.
Going further with the Policy Engine
Open Govern > Policy Engine > MCP Gateway > Controls > Human Confirmation. Confirmation required decides which tool calls need the caller's confirmation and whether a written justification is required.
- Choose a confirmation recipe or start a blank rule under Confirmation required.
- Choose the server, tools and callers under Applies to.
- Set the confirmation outcome, check Reads as, and Save to draft.
- Test the intended caller and an exempt caller in Test > Try a situation, then review and publish.
The highest numeric priority decides. At equal priority, the stricter confirmation value wins. The Don't ask for… recipe creates an explicit exemption, which must have a higher priority than the requirement it overrides. With no matching confirmation rule, an otherwise allowed call runs without asking the caller to confirm.
Match a tool annotation, tag, person, Smart Group, agent, or route instead of maintaining a switch on every server. Invocation & Discovery also offers recipes that allow tools and require confirmation; make sure its access decisions permit the call.
While the engine is on, the governed classic tool switches show Controlled by Policy Engine. Changes through Edit anyway are stored for classic mode; they do not replace the published rule. The client approval experiences below are the same regardless of which mode required confirmation.
What your users see
The approval request appears in the user's AI app. Each app supports different things, so the gateway picks the best option the app tells it it can handle, in this order:
If an in-chat prompt can't be shown or isn't answered, the gateway falls back to an approval link. Cards always come with the Approve in Quilr link underneath, in case the card doesn't load. Card and link resolve the same request, so approving in one place is enough.
Support by client
Based on what each app reports to the gateway when it connects, as of October 2026. Apps change often, so a newer version may behave differently.
Step by step
claude.ai and Claude Desktop
- Claude calls a tool that needs approval. An Approval required card appears in the reply.
- Claude's text written before the card may say the tool "needs approval" or is waiting for you. That is expected: Claude wrote it before you clicked.
- Review the arguments, add a justification if asked, and click Approve. The tool runs once and the card shows Approved - completed with the result.
- claude.ai then places a short message in your message box, starting with Approved, that asks Claude to continue with the result. claude.ai shows its own caution notice above it. Press Enter to send it, and Claude continues from the result.
After you approve, claude.ai shows a red notice above the message box: Use caution before running this prompt. Malicious content could trick Claude into attempting harmful actions or sharing your data. claude.ai shows this for any message an app or card fills in for you, not because anything is wrong with the call. The tool has already run once with the arguments you approved. Check that the message is the short Approved follow-up from the card, then press Enter.
The last step is a claude.ai safeguard: claude.ai doesn't let a card send a message for you without your say-so, and Quilr can't bypass it. The tool has already run at step 3. Sending the message only lets Claude read the result and carry on. claude.ai doesn't support in-chat prompts today, which is why it uses the card.
If the tool takes a long time, the card shows Approved - still running and a Check result button. If you use the Approve in Quilr link under the card instead of the card's own button, the tool doesn't run yet: ask Claude to try the tool again.
ChatGPT
- ChatGPT calls a tool that needs approval. An Approval required card appears in the conversation.
- Click Approve. The tool runs once and the card shows the result.
- The card sends a short follow-up to ChatGPT, and ChatGPT continues with the result without you typing anything.
ChatGPT's text written before the card may still ask you to approve or reply "Approved". That is expected: it was written before you clicked, and ChatGPT carries on once the card reports the result.
VS Code (GitHub Copilot Chat)
- Copilot calls a tool that needs approval. VS Code shows a form in the chat with the tool, the server and the arguments.
- Choose Approve (and type a justification if asked) or Deny.
- On Approve, the waiting call continues and Copilot gets the result straight away.
Claude Code
When Claude Code uses connectors from your claude.ai account, it can't show cards or in-chat prompts, so approval goes through the link:
- Claude Code calls the tool. The reply contains an Approve in Quilr link and tells the assistant that approval is pending.
- Open the link, review the request, and click Approve.
- Claude Code calls the tool again and gets the result. If it tries again before you click, the gateway holds that attempt for up to about 45 seconds so your click turns straight into the result. If you take longer, the assistant gets the link again: approve, then ask it to try again.
Recent versions of Claude Code added directly as an MCP server (rather than through claude.ai connectors) report support for in-chat prompts, so approval appears as a prompt in the terminal or IDE panel instead.
Clients that use the approval link
Copilot Studio, older Claude Code versions and any app without prompts or cards follow the same pattern as Claude Code above: open the link, click Approve, and the tool runs on the assistant's next call. If the assistant doesn't try again on its own, ask it to.
Justification
The justification box reads Justification (required to approve) when the tool requires one and Justification (optional) otherwise. Up to 1,000 characters. Approve without a required justification shows A justification is required to approve. and nothing runs. The justification is saved with the call and is never sent to the MCP server or the model.
Deny, dismiss or no answer
A request that was already approved or denied elsewhere shows Already decided, together with the earlier outcome.
Limits and good to know
- One approval, one run. Each approval covers a single call. Approving on a card or prompt runs the tool once. If the assistant then calls the same tool again, it gets the stored result from that run for up to 10 minutes instead of running the tool a second time, so tools that send, create or delete something don't run a second time just because the assistant retried.
- The approved arguments are what runs. After approving through a link, the assistant's next call of the same tool uses the approval. If the assistant changes the arguments slightly on that retry, the gateway runs the arguments shown on the approval page, not the new ones, as long as that is the only open approval for the tool. Otherwise it asks for a fresh approval.
- Time limits. An unanswered request expires after 10 minutes. An approval given through the link must be used within 5 minutes. A stored result from a card approval is kept for 10 minutes.
- Approval doesn't skip other controls. The approved call still passes through Security guardrails, usage quotas and the other policies in force when it runs. If they block it, it is blocked. Quotas count an approved call once, even if the assistant retries.
- Apps remember tool lists. Most apps cache the list of tools and the card when they connect. After you turn approval on or off for a tool, users may need to reconnect the server, or start a new chat, before the change shows. Until then, the Approve in Quilr link still works.
- Signing in to approve. By default the approval link works without signing in: having the link is enough, and the decision is recorded as the user who made the call. The link is part of the reply, so the assistant can read it too. If you need the approver to be signed in to QuilrAI as that same user, ask QuilrAI to turn on sign-in for approval pages for your deployment (for self-hosted gateways, set
MCP_APPROVAL_PAGE_AUTH=cookie). - The caller must be known. Approval is tied to a person. If the gateway can't tell which user is making a call, calls to tools that need approval are blocked rather than run.
- Unattended agents. Automations such as n8n workflows or scheduled scripts have nobody to click the link, so their calls wait and expire. Leave approval off for those callers with Group and user rules, or match on agent name in the Policy Engine (below).
Review confirmed calls
Open Inspect > Logs on the server card, choose Tool calls, and click a call. Overall analytics opens the separate Insights pages. The Tool call detail drawer shows the arguments and the result. The approval decision, who decided, and the justification are recorded in the call's Metadata.
How this compares to other MCP gateways
Based on each vendor's public documentation as of October 2026. Products change often, so check the vendor's current docs before deciding.
QuilrAI keeps approval in the conversation. In apps that support in-chat prompts or cards, the tool runs the moment the user clicks, without relying on the assistant to try again. With link-based approval, the assistant has to call the tool again after the user approves.
"Not documented" means the vendor's public documentation doesn't describe it, not that the product lacks it.
Sources: MintMCP tool approval, TrueFoundry MCP tool approval, Runlayer agent platforms.
Related
- Tool visibility - enable and disable tools.
- Security guardrails - block or redact sensitive data in calls.
- Policy Engine overview - how cards, stages and priorities work.