withHuman MCP

將您重要的工具呼叫置於人工審核之後。

託管 MCP 伺服器

npx add-mcp 'https://app.withhuman.ai/api/mcp/v1'

可安裝到 Claude Code、Codex、Cursor 等客戶端

文件

The approval layerfor AI agents.

withHuman lets you put your most consequential agent actions behind human approvals, so your agents can do real work.

Framework-agnostic · Open source · Built on open standards

How it works

Follow a consequential action.

A support agent refunds a $75 duplicate charge. Refunds over $20 need a human.

Agentsupport-agent

J

I think I was charged twice this month. Can you check?

Jane · Northwind

Found it: a duplicate $75.00 charge. I'll refund the extra one now.

support-agent

withHuman

Intercept

stripe.refund({ amount: 7500, reason: "duplicate_charge" })

⏸ held by withHuman

Approval pipeline

$75 > $20Skipped

Ask a human for anything moreAsk a human

Escalation path

Payments teamChris BattarbeeOn call

Notify the on-call human from the Payments team.

Notify

SlackDiscordEmailText messageWhatsAppiPhoneAndroid

Chris's notification rules pick the channels.

Decide

stripe.refundfrom support-agent

amount

$75.00

reason

duplicate_charge

customer

[email protected]

Two identical $75.00 charges on Oct 3. The second one is a duplicate.

ApproveDeny

Toolstripe.refund

✓ approved by Chris Battarbee

stripe.refund → re_3QK9rNb2

All done. The $75.00 refund is on its way to your card.

support-agent

Features

Everything you need to let agents do real work.

withHuman covers the full lifecycle, from tool call interception through to human decision.

interception

Every tool call, whatever the agent

Each tool call is intercepted by AAP and sent to withHuman for approval. If the agent supports AAP, withHuman just works.

  • SDKs for LangGraph, CrewAI, Mastra and other agent frameworks
  • Hooks for Claude Code, Codex and other coding agents
  • The open Agent Approval Protocol for your own agents

approval pipelines

No approval fatigue

Stop approval fatigue by writing policy rules that approve or deny routine requests automatically. Only send the tough stuff to humans.

  • Conditions on agent, tool, arguments and agent context
  • An AI judge for calls rules can't describe
  • Webhooks to your own policy service

escalation paths

The right human, every time

Write rules to decide which requests go to which team or person.

  • Route by agent, tool, arguments or team
  • Teams set their own escalation order and on-call flows
  • Escalate when nobody answers in time

destinations

Meet reviewers where they are

Reviewers get requests where they already work, so it fits into their workflow.

  • Slack and Discord
  • iPhone and Android apps with push
  • Email, text message and WhatsApp

audit log

Every decision on the record

Every request, decision and note is kept. When someone asks who approved this, you know.

  • Append-only history
  • Reviewer identity on every decision
  • The full request kept with each decision

Use cases

For the actions you can't take back.

Let agents draft everything and ship nothing until someone accountable has looked.

# approvalsSlack

withHumanAPP9:11 AM

sre-agent requests approval for kubernetes.rollout_undo

Roll payments-api back to v2.41.0 after the 5xx rate crossed 2%.

image: v2.42.1 v2.41.0

ApproveDeny

Integration

Works with the agents you already use.

Install the CLI to connect every supported agent on your machine. For your own agents, use the AAP SDKs for LangGraph, CrewAI, Mastra and more.

$curl -fsSL https://downloads.withhuman.ai/install.sh | sh -s -- onboard

# Connects Claude Code, Codex,
# OpenClaw, Pi, Hermes and
# DeepSeek on this machine.

Hooks are in place.

Finish in the browser: the request
appears there for you to approve
or deny.

Open standards

No lock-in. It's a protocol.

Your agents speak to withHuman via AAP, the Agent Approval Protocol. AAP is an open spec started by withHuman to allow agents to wait on human decisions. Your integrations are with AAP, not withHuman.

RFC

Built in the open

The spec evolves in public: issues, proposals, and changes all happen where you can see and engage with them.

aap/request

One integration, any provider

Agents and frameworks integrate against AAP, not against withHuman. We aim to be the best provider, never the only one.

open source

Leave whenever

The spec, SDKs, adapters, and conformance tests are open source. Swap providers without rewriting a single agent.

withhuman/core

Self-hostable

The withHuman core is open source: the queue, the audit trail, and the web UI. Run it on your own infrastructure or let us run it for you.