Human in the Loop
HITL logowithHuman logo

HITL vs withHuman

HITL lets your agent ask a person anything, not only whether a tool call may run: get a number, a rating or an edited draft back, chain several reviews, and keep going with a default answer when nobody replies.

Branching workflowsLast updated Sources listed below

TL;DR: three reasons teams pick HITL

HITL keeps a person in the loop on AI agents, from a one-tap approval to a multi-stage review that branches on every answer. Here is where it pulls ahead of withHuman.

  • More than approve or deny: 7 response types, including editable text, a number and a rating, so the agent gets real input back.

  • Several Human Review steps in one workflow, each branch with its own reviewers, where withHuman ends a request at the first decision.

  • A default answer on timeout, Telegram, an Apple Watch app and n8n, Zapier and Make, with flat pricing from $10 a month.

Built for complex decisions, not just yes or no

In HITL a workflow is a graph, not a straight line. Classify the request, ask the right person, branch on what they say and merge back. Build it on a canvas, or describe it and let the workflow assistant build it for you.

Live in the HITL portalSeveral Human Review steps per workflow
New support emailTrigger · Email
Sort the requestAI Classifier
Refund
Approve the refund?Human Review · Finance
Approved?Condition

Yes: Send the refund · No: Email the customer

Escalate
Lead decidesHuman Review · Support lead
Post to #supportAction · Slack
Other
Draft a replyAI
Edit the draftHuman Review · Editable text
Email the customerAction · Email
EndEvery path finishes here
A support workflow in HITL. An AI Classifier sorts each email: refunds go to Finance and branch on their answer, escalations go to a lead, everything else gets an AI draft that a person edits. Three Human Review steps, one workflow.
  • As many reviews as the decision needs

    Put several Human Review steps in one workflow: a first check, a second sign-off, an escalation, and a different reviewer on each branch.

  • Branch on the answer

    A Condition step routes on what the person decided or on the data. An AI Classifier sends each request down the path for its category.

  • Paths that meet again

    Branches can merge back into one path, so a refund, an escalation and a drafted reply can all finish with the same follow-up step.

  • Describe it, get the workflow

    The built-in workflow assistant builds the graph from a plain-language prompt. Every revision is kept, so you can restore an earlier one.

Nine step types on the canvas

  • TriggerStart on a schedule, an email or a spreadsheet row.
  • Human ReviewAsk a person to decide, with any of the 7 response types.
  • AIRun a prompt on the data so far.
  • AI ClassifierSort input into categories, each with its own path, plus Other.
  • ConditionBranch yes or no on the answer or on any data.
  • TransformSet variables for the steps that follow.
  • DataFetch from Google Sheets, an API or a table.
  • ActionCall a webhook, send an email or post to Slack.
  • EndClose the run once every path is done.

What about withHuman?

withHuman's approval pipelines check a tool call against Branch, Delegate, AI judge and webhook blocks, then approve it, deny it or ask a person, and an escalation path decides who is contacted and when. Its docs say one approval or denial ends the request, with no requirement for several sign-offs. HITL chains Human Review steps and branches on each answer.

Feature comparison

Where HITL leads comes first. Then the features HITL and withHuman share, from connecting agents to pricing.

Feature comparison of HITL and withHuman. HITL's advantages are listed first.
FeatureHITLwithHuman
Nonlinear workflows (branching, multi-step review)YesBranches, AI routing, several review stepsNoBranching rules, then one decision
Callback webhooksYesPer requestNot listedAgents wait on a long poll
No-code tools (n8n, Zapier, Make)Yesn8n, Zapier, MakeNot listed
Answers beyond approve or denyYes7 response typesNoApprove or deny, with a note
Reviewer edits the AI draftYesEditable textNot listed
Timeouts with a default answerYesNoRequest expires; the agent must not act
Apple WatchYesNot listed
TelegramYesNot listed
Visual workflow editorYesCanvas with 9 step typesYesPipeline of blocks
MCP server for agentsYesHosted, with sign-inYesMCP gateway
REST APIYesYesAgent Approval Protocol
One-step setup for coding agentsYesPaste-in setup promptYesCLI guided setup
Web inboxYesmy.hitl.shYesReview queue
iPhone appYesYesScale plan, $20 per user
Android appYesYesScale plan, $20 per user
Free planYes500 requests / moYesUp to 5 members
SlackComing soonYesAlso Discord, SMS, WhatsApp

"Not listed" means we could not find it on withHuman's public site or docs in October 2026. It may still exist. "Coming soon" channels are on HITL's roadmap and not live yet.

The products, side by side

HITL's web inbox next to a product image withHuman shows on its own site.

HITL web inbox with a refund request open, showing the customer's message and Approve refund and Deny buttons
HITL web inboxfrom hitl.sh
withHuman review queue with an issue_refund tool call open, showing its arguments, the agent's explanation and Deny and Approve buttons
withHumanfrom withhuman.ai

About withHuman

withHuman is an open source approval layer for AI agents, backed by Y Combinator. It intercepts the tool calls an agent makes through hooks, SDKs or an MCP gateway, runs each one through an approval pipeline of rules, and sends the ones that need a person to a reviewer, who approves or denies it. Reviewers are reached by email, Slack, Discord, text message, WhatsApp and apps for iPhone, Android and Mac.

Why teams pick HITL

For the people who answer the requests and the teams who run the workflows.

  • Ask for input, not only permission

    Have a person pick a reply, fix a draft, enter an amount or score an output. The agent gets typed data back, not just a yes or no.

  • Several sign-offs when it matters

    A first reviewer checks, a Condition step reads the answer and the big cases go to a second reviewer. Each branch can have its own people.

  • Work keeps moving

    Every request carries a default response. If nobody answers in time, the agent gets that answer and carries on instead of stopping.

  • Flat monthly pricing

    Free for 500 requests a month and 3 humans, then a flat $10 or $30 a month, instead of a price per reviewer.

Why agents work well with HITL

For the agents asking the questions. HITL gives them a clear contract: ask, wait, get a typed answer.

  • A hosted MCP server

    Add https://mcp.hitl.sh/mcp to Claude Code, Cursor or any client that supports remote MCP. Sign in once and the agent gets tools like list_loops, create_request and get_request.

  • A structured answer, every time

    The agent picks one of 7 response types and gets back typed data: a decision, the chosen options, a number, a rating or the edited text. No parsing a chat reply.

  • Timeouts that keep work moving

    Every request carries a default response. If nobody answers in time, the agent gets that answer and carries on instead of waiting forever.

  • A setup prompt the agent follows

    Copy the setup prompt from hitl.sh and paste it into your agent. It asks for your API key, reads the docs at docs.hitl.sh/llms.txt and lists your loops to prove the connection works.

Add HITL to Claude Code

claude mcp add --transport http hitl https://mcp.hitl.sh/mcp

Questions

What teams ask when they weigh HITL against withHuman.

Why choose HITL over withHuman?
withHuman decides whether an agent's tool call may run. HITL does that too, and also lets the agent ask a person for a choice, text, an edited draft, a number or a rating, chain several reviews in one workflow and fall back to a default answer on timeout.
Can HITL run multi-stage approvals?
Yes. A HITL workflow can hold several Human Review steps, branch on each answer and give each branch its own reviewers. withHuman's docs say one approval or denial ends a request, and its escalation path contacts more people over time rather than asking for more sign-offs.
What happens when nobody answers?
In HITL the agent gets the default response you set and carries on. In withHuman the request expires and the agent must not carry out the action.
Do both work with coding agents?
Yes. withHuman installs hooks for Claude Code, Codex and other coding agents with its CLI. HITL connects through its hosted MCP server, added to Claude Code with one command, or a setup prompt you paste into Claude Code, Cursor or Codex.
How does pricing compare?
HITL is free for 500 requests a month and 3 humans, then a flat $10 or $30 a month. withHuman is free for up to 5 members, then $20 per user per month on its Scale plan, which is also where its phone apps and text alerts start. Its open source edition is free to self-host. HITL includes its mobile app on the free plan.
Can I use both?
Yes. withHuman can gate which tool calls run, and HITL can ask a person for the input an agent needs, with richer answers and a branching workflow.

Ship AI workflows your team can trust

Free to start. Your first workflow takes minutes.

Get started free