Set up human approval for your AI agents' tool calls.
ai.withhuman/withhuman MCP Server
The ai.withhuman/withhuman Model Context Protocol (MCP) server enables setting up human approval for AI agents’ tool calls. Its purpose is to introduce a review/approval step before agent tools are executed, based on the server description provided.
🛠️ Key Features
Human approval for AI agents’ tool calls
🚀 Use Cases
Require human confirmation before an agent triggers tool calls
⚡ Developer Benefits
Adds an explicit approval gate to agent tool execution
⚠️ Limitations
Available source data only specifies human approval for tool calls; no other capabilities, tools, or configuration details are provided.
We want agents to do more than read data and suggest what to do next. They
should be able to roll back a deployment, issue a refund, or merge code.
Those actions have consequences, and teams need a way to bring a person in
before they happen.
withHuman is the approval layer for that moment. It checks an agent's tool
call before it runs against rules you define in one place. Routine calls
can be approved automatically, disallowed calls can be denied, and
consequential ones can wait for the right person to decide. Every agent you
connect follows those rules.
How it works
An adapter (a runtime hook or an SDK wrapper) sends each tool call to
withHuman before it runs, over the open
Agent Approval Protocol.
sequenceDiagram
participant Agent as Your agent
participant WH as withHuman
participant You as You
Agent->>WH: Wants to call issue_refund(amount: 75)
WH->>WH: Checks your approval pipeline
WH->>You: Asks for approval by email, Slack or Discord
You->>WH: Approve
WH->>Agent: Go ahead
Agent->>Agent: Runs issue_refund
Approval pipelines approve, deny or ask a person, based on the tool
and its arguments. For example: approve refunds up to $20, ask payments
above that.
Escalation paths choose who is asked and when, by team, level and
working hours. Reviewers get email, Slack or Discord and decide in the
web app with the exact arguments in front of them.
Open http://localhost:8080, choose First
setup and enter the secret. The welcome page connects your first agent
and ends with a test call for you to approve. The
quickstart guide
covers each step.
To run from source instead, install Go (the version in go.mod)
and Node 25, then run make frontend && go run ./cmd/withhuman. The
secret is in ~/.withhuman/bootstrap-secret.
Connect your agents
Coding agents. Claude Code, Codex, OpenClaw, Pi, Hermes and DeepSeek
connect with the withhuman CLI:
sh
curl -fsSL https://downloads.withhuman.ai/install.sh | sh
withhuman agent install --url https://approvals.example.com
Agents you build. LangGraph, Pydantic AI, the OpenAI Agents SDK,
CrewAI, Mastra and other frameworks connect through the
AAP SDKs. See the
examples.
Settings are environment variables, listed in .env.example.
Before you go beyond your own machine:
Serve it behind HTTPS and set WITHHUMAN_PUBLIC_ORIGIN to that origin.
Set WITHHUMAN_TRUSTED_PROXIES to your proxy's addresses only.
Set WITHHUMAN_SECRETS_KEY and keep it safe.
Configure SMTP so invitations and password resets reach people.
Back up the whole data directory.
Set OTEL_EXPORTER_OTLP_ENDPOINT to collect its OpenTelemetry metrics.
Don't want to run it yourself? withhuman.ai hosts
it and adds the audit log, the AI judge, the MCP gateway, phone
notifications and single sign-on.
Telemetry
Once a day withHuman sends a random installation ID, its version, the
platform and whether it runs as a binary, a container or on Kubernetes.
The answer tells administrators about new releases and security fixes on
the Version settings page. Nothing about your organizations, people,
agents or requests is sent. Set WITHHUMAN_TELEMETRY=off (or
DO_NOT_TRACK=1) to turn it off. See
Updates and telemetry.