human-in-the-loop · sub 60s · api-first

your agent can do everything.
except log in.

it passed the bar. it writes code, drafts contracts, manages your life. it cannot solve a captcha designed for a 6-year-old.

get api key →become an operator

5 free relays/month · no credit card · yes, it's real

38s
median relay time
$0.75
per authentication
0
captchas solved by AI
the problem
what your agent can do
  • ✓ write, debug, and deploy code
  • ✓ research 200 sources in 4 minutes
  • ✓ draft emails your team would send
  • ✓ pass the bar exam, USMLE, and CPA
  • ✓ reason over 200k tokens of context
what stops it cold
  • ✗ "select all images with traffic lights"
  • ✗ "enter the code we texted you"
  • ✗ "sign in with google" (real session required)
  • ✗ okta / ping / saml sso
  • ✗ email magic link (no inbox)
authrelay plugs humans into the gap. agent calls us → human logs in → agent continues. that's it.
what we handle
🤖
captcha walls
cloudflare, recaptcha, hcaptcha. your agent aced the bar exam. a fire hydrant grid has it stumped.
🔑
oauth & sso logins
"sign in with google" requires a real session. your agent has neither cookies nor shame.
📱
2fa & sms codes
the code goes to your phone. your agent doesn't have a phone. this is still somehow a strategy in 2026.
🏦
bank & fintech auth
"enter the 3rd, 7th, and 9th character of your security word." designed to stop bots. it works.
🧑‍💼
corporate sso
okta, ping, or some bespoke auth system from 2009. we've seen them all. yes, including that one.
📧
email magic links
"click the link we sent you." your agent can't check email. a human can. link clicked. done.
🤖 for agents
native mcp integration
claude, gpt, cursor — one tool call:
request_human_auth
view setup →
how it works
1
agent hits an auth wall
captcha. login page. 2fa prompt. instead of failing silently or hallucinating credentials, your agent calls our api in one line.
2
operator gets the ping
a verified human is notified. they accept and get a scoped, time-limited browser window — just the login page. nothing else is visible.
3
human does the thing
logs in. solves the captcha. enters the code. clicks "i agree to terms" on your behalf. session access expires the moment they're done.
4
agent continues
status flips to "completed". agent gets the signal and picks up exactly where it left off. elapsed time: ~38 seconds.
in the wild

real agents. real bottlenecks. real resigned acceptance.

s
swyx
@swyx

my agent autonomously researched, drafted, and sent a report to 50 people. then it got stuck on a linkedin login screen for 4 hours. authrelay fixed it. i feel nothing.

9:14 AM · Mar 3, 2026♥ 3.2K
l
levels.io
@levelsio

had an agent doing competitive research across 12 sites. 4 had captchas. $6 in authrelay credits later, the agent finished the job. this is the future and it's extremely dumb and i love it

11:42 PM · Feb 28, 2026♥ 8.7K
A
Andrej Karpathy
@karpathy

the funniest part of the AI agent era is that the biggest bottleneck is often a checkbox that says "i'm not a robot" trained on 100 trillion tokens. a 6x6 grid of traffic lights has us beat.

2:31 PM · Feb 19, 2026♥ 41K

* illustrative. the problem is extremely real.

one import. done.

your agent doesn't need to know how it works. it just needs to know it does.

⚡ agent self-signup — one curl, zero browsers
curl -X POST https://authrelay.app/api/auth/register \
  -H "Content-Type: application/json" \
  -d '{"email":"[email protected]","password":"securepass123"}'

# → { "api_key": "ar_live_...", "quickstart": { ... } }
python
from authrelay import AuthRelay

relay = AuthRelay(api_key="ar_live_...")

result = relay.request_auth(
  url="https://example.com/login",
  reason="oauth required"
)

# ~38s later
cookies = result.credentials["cookies"]
# ✓ agent continues with auth cookies
node.js
import { AuthRelay } from '@authrelay/sdk'

const relay = new AuthRelay({
  apiKey: 'ar_live_...'
})

const result = await relay.requestAuth({
  url: 'https://example.com/login',
  reason: 'oauth required'
})

// credentials.cookies → inject into browser
// ✓ agent continues authenticated
pricing

you pay per login. not per month of forgetting to cancel.

pay-per-use
$0.75/relay

only pay when it works

  • ✓ flat $0.75 per authentication
  • ✓ no time-based charges
  • ✓ node + python sdk
  • ✓ mcp integration
  • ✓ full audit log
  • ✓ encrypted credential handoff
get started
enterprise
custom

for teams with volume

  • ✓ dedicated operator pool
  • ✓ custom sla
  • ✓ priority queue
  • ✓ volume discounts
  • ✓ soc 2 compliance
contact us

no monthly fees. no caps. just pay when a human saves your agent. enterprise inquiry →

faq
is this real?
yes. your agent calls an api. a real human gets a notification, takes over a time-limited browser window, and logs in. access is revoked the moment they're done. it's extremely normal and definitely fine.
is the human seeing my data?
only the login screen — same as handing your laptop to IT support. operators are verified, NDA-bound, and the browser session is scoped to just the auth page. full audit log for everything.
how fast is "under 60 seconds"?
median is 38 seconds from request to completion. if no one's available within 10 minutes, session expires and your agent can retry automatically.
why not just train a model to solve captchas?
captcha companies train specifically against that. you'd be funding an arms race. alternatively: one human, $0.75, 38 seconds. up to you.
can my agent become an operator?
no. that would defeat the entire purpose. the robots are the clients here.
what if i have a ton of logins?
pay-per-use scales infinitely — no monthly caps. enterprise gets a dedicated operator pool, volume discounts, and custom sla. most workflows need auth once per domain, so costs stay low.
🔒
access expires on completion
no lingering sessions
✅
verified operators
id-checked + nda-signed
📋
full audit log
every session recorded
⚡
api-first
agent hires autonomously

unblock your agent.

it trained on the entire internet.

it deserves better than dying on a login page.

get api key →

want to earn money being a human? become an operator →