# withHuman

> Approval layer for AI agents that holds consequential tool calls for rule-based or human approval before they run.

withHuman, made by Metoro Inc., is an approval layer for AI agents. It intercepts an agent's tool calls, checks them against approval rules defined in one place, and either approves, denies, or sends them to a human reviewer before they execute. It is available as a hosted service and as a self-hostable core distributed on GitHub under the Elastic License 2.0.

## What It Is

withHuman sits between an agent and the tools it calls. An adapter (a runtime hook or SDK wrapper) sends each tool call to withHuman over the open Agent Approval Protocol (AAP). Rules in an approval pipeline can approve or deny routine calls automatically, for example approving refunds under a set amount, and pass the rest to people. Reviewers see the exact tool call and its arguments and choose to approve or deny; the agent then proceeds or is blocked.

## How Requests Are Routed

- Approval pipelines evaluate conditions on agent, tool, arguments and agent context. The docs also list AI judge blocks, webhook blocks, branch blocks and delegate blocks.
- Escalation paths decide who is asked and when, by team, level and working hours, and can escalate if nobody answers in time.
- Reviewers are notified via Slack, Discord, email, text message, WhatsApp, or the iPhone and Android apps.
- An append-only audit log keeps each request, decision, note and reviewer identity.

## Connecting Agents

The withhuman CLI connects coding agents such as Claude Code, Codex, OpenClaw, Pi, Hermes and DeepSeek. An MCP gateway can keep MCP server credentials in withHuman and route every call through the pipeline. For agents built in code, AAP SDKs cover LangGraph, Deep Agents, Pydantic AI, OpenAI Agents, Claude Agent SDK, Google ADK, CrewAI, AI SDK and Mastra, plus an HTTP API.

## Open Standard and Deployment

The vendor describes AAP as an open spec started by withHuman, with open-source SDKs, adapters and conformance tests, so integrations target the protocol rather than withHuman specifically. The core (queue, audit trail, web UI) can be run with a Docker image; the README says the hosted service adds the audit log, AI judge, MCP gateway, phone notifications and single sign-on.

## Features
- Interception of agent tool calls before execution
- Approval pipelines with conditions, AI judge, webhook, branch and delegate blocks
- Escalation paths by team, level and working hours
- Notifications via Slack, Discord, email, text message, WhatsApp and mobile apps
- Append-only audit log with reviewer identity
- MCP gateway for downstream MCP servers
- Open Agent Approval Protocol with SDKs
- Self-hostable open-source core
- Single sign-on and directory sync
- HTTP API and OpenAPI schema

## Integrations
Claude Code, Codex, OpenClaw, Pi, Hermes, DeepSeek, LangGraph, Deep Agents, Pydantic AI, OpenAI Agents SDK, Claude Agent SDK, Google ADK, CrewAI, AI SDK, Mastra, Slack, Discord, WhatsApp, MCP

## Platforms
WINDOWS, ANDROID, IOS, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://withhuman.ai
- Documentation: https://withhuman.ai/docs
- Repository: https://github.com/withHumanAI/withHuman
- EveryDev.ai: https://www.everydev.ai/tools/withhuman