HumanLayer alternatives for human approval of AI agents
HumanLayer removed its approval SDK in 2025 and is joining Factory. Six ways to keep agents asking a person first, with prices, drawbacks and how to move a require_approval decorator.
Facts checked . Written by the people who make VariantGate, so read our section with that in mind; every claim about another product links to its own pages.
The short answer
- VariantGate if a team’s agents should ask in one inbox, with rules for the routine, amount limits and a record of every decision.
- gotoHuman if people mostly review and edit what an AI wrote before it goes out.
- hitl.sh if the point is reaching a person on their phone, watch or Telegram.
- Permit.io if you already use it for authorization in your own product.
- Your framework's pause if your app’s own user approves in your own screen.
What happened to HumanLayer’s approvals
HumanLayer started as an API and SDK for agents to ask people for approval, best known for the @hl.require_approval() decorator and hl.human_as_tool(). Since then:
- June 2025: the last SDK releases, 0.7.9 on PyPI and npm.
- September 2025: the SDKs were removed from the repository, with the note “this is codelayer now”. The old approval docs no longer load.
- Today: HumanLayer is a control plane for coding agents such as Claude Code and Codex, and on 4 November 2026 Factory announced that HumanLayer is joining it.
If you use HumanLayer to run coding agents, this page is not about that product. It is for people who relied on, or were about to adopt, the approval SDK. We could not confirm whether the old SDK’s cloud API still answers; with the code removed and the docs gone, we would not start anything new on it.
How we compared
Five questions for each option: what it is for, where the person answers, whether rules can answer routine requests in advance, where it runs, and the entry price. Facts about other products come from their own websites, docs and pricing pages on 18 November 2026, listed under Sources; we did not test them hands-on. Facts about VariantGate come from our own product.
| Option | What it is | Where a person answers | Rules decided in advance | Hosting | Starts at |
|---|---|---|---|---|---|
| VariantGate | What it isAn approval inbox and rules for a team's agents | Where a person answersWeb inbox, email, push, Slack, Teams, Discord | Rules decided in advanceAllow, ask and deny by action and agent key, amount limits, budgets (Pro) | HostingHosted in Germany; self-hosting on Enterprise | Starts atFree: 1 person, 100 requests a month. Pro €9 a month |
| gotoHuman | What it isReviews of AI work, with forms to edit it | Where a person answersWeb inbox, email, Slack | Rules decided in advanceNot described on its public pages | HostingHosted | Starts atFree: 1 user, 300 reviews a month. Team $79 a month |
| hitl.sh | What it isA way for agents to reach people | Where a person answersIts own app, Slack, Telegram | Rules decided in advanceNot described on its public pages | HostingHosted | Starts atFree: 3 people, 500 requests a month. Starter $10 a month |
| Permit.io Access Request MCP | What it isAccess and approval requests on Permit's authorization | Where a person answersElements you embed in your own app, or its API | Rules decided in advancePermit's own policies | HostingYou run the open-source MCP server; it calls the Permit API | Starts atCommunity: free to 1,000 monthly users |
| Your framework's own pause | What it isA pause in LangGraph, the OpenAI Agents SDK, the Vercel AI SDK or Cloudflare Agents | Where a person answersWhatever screen you build | Rules decided in advancePer tool, in code | HostingYour own app | Starts atPart of the library |
| A workflow engine | What it isA durable wait in Inngest or Temporal | Where a person answersWhatever screen you build | Rules decided in advanceIn your code | HostingTheir cloud or your own | Starts atSee their pricing |
Prices are the lowest paid plan or the free plan, as listed on 18 November 2026. US dollar prices are as the vendor shows them.
VariantGate
Our product. An agent asks with one HTTP call, an MCP tool call, a hook or the Python and TypeScript clients, and acts only when the answer says allowed: true. People decide in a web inbox, from an email, a push notification or a Slack, Teams or Discord message. Standing rules allow, ask or deny by action and agent key, with amount limits and budgets that fail toward asking. Every decision lands on a sealed record, and one switch pauses every agent.
- Best for
- A team whose agents (in code, in n8n, or Claude Code and Codex on laptops) should all ask in one place, with rules for the routine and a record of every decision.
- The catch
- It answers yes or no, with a note and, on Pro, an edited amount or wording. It has no equivalent of
human_as_toolfor an agent asking a person an open question. It is also new: public since November 2026. - Price
- Free for 1 person and 100 requests a month; rules and budgets from Pro at €9 a month; teams from €15 per person a month. See pricing.
gotoHuman
A hosted inbox for human reviews of AI work. You define review forms without code, reviewers can edit the AI’s output in place, and the finished review is sent to your workflow by webhook so it can carry on. It has an n8n node, a Make integration, TypeScript and Python SDKs, a LangGraph example and an MCP server. We compare it with VariantGate in more detail in VariantGate vs gotoHuman.
- Best for
- Workflows where a person reviews and corrects what an AI wrote (an email draft, a post, extracted data) before it goes out, especially in n8n or Make.
- The catch
- Its public pages describe reviews and forms, not rules that answer routine requests in advance or limits on amounts. It is hosted only, with no self-hosting described.
- Price
- Free for 1 user and up to 300 reviews a month; Team $79 a month with 5 users included (excluding VAT).
hitl.sh
A hosted service built around agents reaching people: its own iPhone and Watch apps, Slack and Telegram, with an MCP server, a REST API and nodes for n8n and Zapier. Of the options here it is closest in spirit to HumanLayer’s idea of a human as something an agent can contact.
- Best for
- Reaching a person wherever they are, including a phone app, a watch or Telegram, when an agent needs an answer.
- The catch
- Its site says the app and Telegram are live today; check that the channel you need is available before you rely on it. Plans are capped by requests per month.
- Price
- Free for 3 people and 500 requests a month; Starter $10 a month for 2,000; Pro $30 a month for 5,000 (USD).
Permit.io Access Request MCP
An open-source MCP server from Permit.io that gives an agent tools to ask for a role on a resource, or for one-time approval of an action, and leaves the decision to a person. Requests and approvals go through the Permit API and its policies.
- Best for
- Products already using Permit.io for authorization, where an agent asking for access or for one action is another kind of permission request.
- The catch
- You run the MCP server and show approvals in Permit's elements inside your own app, so the reviewer's screen is yours to build and host.
- Price
- Permit's Community plan is free for 1,000 monthly active users and 20 tenants; Enterprise on request.
Your framework's own pause
LangGraph’s interrupt, the OpenAI Agents SDK’s needs_approval, the Vercel AI SDK’s tool approval and Cloudflare Agents’ waitForApproval() all stop a run before a tool and resume it with a person’s answer. With a durable checkpointer they survive a restart, which no outside service does for you. Human-in-the-loop compared goes through them one by one.
- Best for
- One application where the person deciding is that app's own user, in a screen you already have.
- The catch
- The pause is all you get. Notifications, the reviewer's screen, who approved what, and any rules beyond a per-tool switch are yours to build.
- Price
- Part of the library.
A workflow engine
Inngest’s step.waitForEvent() pauses a run until a matching event arrives or the wait times out; Temporal workflows wait for a signal. Both hold the state of a long process far better than a polling loop. Pair one with any of the services above, or build the rest yourself.
- Best for
- Teams already running Inngest or Temporal, where an approval is one more step in a long-running workflow.
- The catch
- They give you a durable wait, not an approval product: the message to the person, the screen, the record and the rules are still yours.
- Price
- See Inngest's and Temporal's own pricing.
Moving a require_approval decorator
The pattern carries over to any of the services above: wrap the risky function, ask before it runs, run it only on a yes. With VariantGate’s Python client, a HumanLayer function like this:
from humanlayer import HumanLayer
hl = HumanLayer()
@hl.require_approval()
def send_email(to: str, subject: str, body: str) -> str:
...becomes this:
from variantgate import ApprovalClient
approvals = ApprovalClient("mail-agent") # reads VARIANTGATE_KEY
@approvals.require_approval(
"email.send",
lambda to, subject, body: f"Email {to}: {subject}",
details=lambda to, subject, body: {"to": to, "body": body},
)
def send_email(to: str, subject: str, body: str) -> str:
...Differences worth knowing before you switch:
- You name the action (
email.send) and write the summary the approver reads, so rules can match the action and the inbox stays readable. - A denial, a timeout and any error raise
NotApprovedErroror an API error, so the function never runs without a yes. Catch it where your agent should hear “not allowed”. - There is no
human_as_tool. If your agent asks people open questions, look at gotoHuman’s forms or hitl.sh for that part.
Framework guides: LangGraph, CrewAI, OpenAI Agents SDK, Pydantic AI and the rest.
Sources
- HumanLayer pull request #646, merged 29 September 2025: “Clean up humanlayer sdk stuff, this is codelayer now”, which removed the SDKs.
- PyPI: humanlayer 0.7.9, released 3 June 2025, the last Python SDK release.
- npm: @humanlayer/sdk 0.7.9, released 3 June 2025, the last TypeScript SDK release.
- HumanLayer today: a control plane for coding agents, Starter free up to 3 members and 200 sessions a month, Pro $100 per user a month.
- Factory, 4 November 2026: HumanLayer is joining Factory; the current platform stays up during a transition period.
- gotoHuman: custom human reviews to approve and edit any action; n8n node, Make, TypeScript and Python SDKs, LangGraph example, MCP server; completed reviews sent by webhook.
- gotoHuman pricing: Free $0 for 1 user and up to 300 reviews a month; Team $79 a month with 5 users; prices excluding VAT.
- hitl.sh: agents reach people through its app, Slack and Telegram; MCP server, REST API, n8n and Zapier nodes.
- hitl.sh pricing (USD): Free 500 requests a month for 3 people; Starter $10 a month for 2,000; Pro $30 a month for 5,000.
- Permit.io: Access Request MCP, an open-source server that lets an agent ask for access or for approval of a sensitive action.
- Permit.io pricing: Community free for 1,000 monthly active users and 20 tenants; Enterprise on request.
- LangGraph: Interrupts. Pause a graph and wait for outside input.
- Vercel AI SDK: tool approval (toolApproval; needsApproval is deprecated), for tools the SDK runs itself.
- Cloudflare Agents: human in the loop. waitForApproval() as a durable gate on Cloudflare Workflows.
- Inngest: step.waitForEvent() pauses a run until a matching event arrives or the wait times out.