NobleCloak
Human in the loop

Not just a human in the loop. The right one.

Every agent framework can pause and ask for approval. None of them knows who should answer. NobleCloak sends each risky action to the person your company says should decide, and keeps the record.

What human in the loop means

The agent proposes. A person decides.

Human in the loop means an AI agent stops before a risky action and waits for a person to approve or deny it. The agent carries on with the answer. The pause is the easy part: LangGraph, Claude Code and most agent SDKs already have one. The hard part is who answers it.

The wrong humans

Built-in approval asks whoever is running the agent.

That's usually the wrong person.

The person running the agent

Built-in approval asks them by default, whatever the action. With NobleCloak, the request goes to the approver your policy names for that action.

The developer

They wrote the agent but have no business context for a vendor payment or a refund. With NobleCloak, the request goes to the business owner, by name or by role.

Whoever is online

They can click yes but don't have the authority to decide. With NobleCloak, approvers come from roles your organization defines.

Nobody

After a week of prompts, everyone switches to always-allow. With NobleCloak, only the actions your policy flags reach a person. Everything else runs or is refused without asking.

The parent session

A sub-agent acts on approval it was never given. With NobleCloak, one approval covers one action, and sub-agents never inherit it.

one policy, any runtime
ActionWhoDecision
release_paymentAP agentneeds approval
delete_production_tableany agentneeds approval
issue_refundsupport agentneeds approval
edit_vendor_bankany agentdenied

Payments go to the AP manager. Production deletes go to the engineering lead. Refunds go to the support lead.

Illustrative: names no package, import or API call

How to add human approval to an AI agent

Five steps, in order.

1. Install the plugin for your runtime

LangGraph (TypeScript), Claude Code, or any MCP client through the NobleCloak gateway.

2. Mark the tools that need a person

Each tool is allowed, needs approval, or is denied.

3. Name who decides

A person or a role, such as Finance or the engineering lead. Naming approvers starts on the Business plan.

4. The agent pauses and the approver answers

In Slack or the web inbox. Slack delivery starts on the Team plan; Teams and email are coming soon. The agent resumes exactly once.

5. Every decision is recorded

The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when.

Your framework already pauses

The plugin hooks the pause your runtime already has.

Then it asks NobleCloak who decides and hands back the answer.

  • LangGraph

    TypeScript · Python coming soon

    Available
  • Claude Agent SDKat launch
  • Claude CodeAvailable
  • Any MCP client

    Through the gateway

    Available
  • Vercel AI SDKComing soon
  • OpenAI Agents SDKComing soon
  • Strands / AgentCoreComing soon
  • Google ADKComing soon
Questions

Human in the loop, answered

What is human in the loop for AI agents?

An AI agent stops before a risky action and waits for a person to approve or deny it. NobleCloak decides which person: the one your policy names for that action.

How do I require approval for AI agent actions?

Install the NobleCloak plugin for your runtime, set the tools that need approval in policy, and name who approves each one. The agent pauses on those tools and carries on when the approver answers.

Does LangGraph support human in the loop?

Yes. LangGraph can interrupt a run on its checkpoint and resume it later. The NobleCloak plugin uses that pause and adds who decides, plus the record. It's available for LangGraph in TypeScript today; Python isn't published yet.

How is this different from my framework's built-in approval?

Built-in approval asks whoever is running the agent, in that session. NobleCloak routes the request to the person your policy names, wherever they work, applies one approval to one action, and records every decision across every runtime.

Does every tool call need a person?

No. Each tool is allowed, needs approval, or is denied. Only the tools your policy marks for approval reach a person, which is what keeps approvals from becoming a rubber stamp.

What does it cost?

The Free plan is $0. Team adds sharing across ten users with Slack delivery (Teams coming soon). Business adds named approvers, role-based policies and audit export. See the pricing page for the full comparison.

Put the right person in the loop.

Start with the plugin for the runtime you already use.