Contents
- The problem
- What Login Direct is
- How a job works, step by step
- Features
- What it will never do
- Strengths
- Limitations
- Where it works best
- Using it in the web app
- Using it as an MCP server (Claude and other AI clients)
- Security and privacy
- FAQ
- Responsible use
The problem
Think about how many times you have done this:
- You want to read an article, try a tool, download a file or join an event.
- The site asks you to create an account.
- You type a name, an email, a password. You confirm a verification email. You solve a CAPTCHA.
- A week later your inbox is full of marketing from a site you used once.
For one site it's a minor annoyance. For a developer testing sign-up flows, a researcher who needs to look at several sites, or anyone who prefers not to hand a personal email to every service, it adds up.
There is a second problem on the other side. AI assistants are good at reading and reasoning, but they can't use a site that sits behind a login. The assistant can't sign up, so it can't see what's inside.
Login Direct deals with both.
What Login Direct is
Login Direct is a web application with an AI agent at its core. You give it a link. The agent opens a real browser, looks at the page, fills in the form and completes the sign-up, then reports back.
It can work in two modes:
- Temporary accounts (the default). The agent invents a throwaway identity: a name, a username, a strong random password and a disposable email address. It signs up, reads the verification email by itself, and saves the credentials in your private vault. Your own email is never involved.
- Permanent accounts. If you want an account you'll keep, you connect your own email and give the agent your saved details. It then signs up with your information and reads verification messages from your inbox.
It can also fill in and submit online forms (Google Forms, Microsoft Forms, Typeform, Jotform, Tally and others), and it can be driven by other AI tools through MCP, which means Claude can ask it to register on a site and bring back what it finds.
How a job works, step by step
Every request you make is a job. A job moves through these stages:
Queued → Starting → Opening site → Analyzing → Identity created → Filling form → Submitting → Waiting for email → Verifying → Completed
At almost any point it can also branch to Human action required (a CAPTCHA or another check), Waiting for your answer (the agent needs a detail from you), or end as Failed or Cancelled.
Here is what happens behind the scenes:
- You paste a URL. The backend first checks that the address is safe. Private and internal addresses are refused, so the tool can't be pointed at someone's internal network.
- A background worker takes the job. Browser work can take minutes, so it never runs inside a web request.
- A real browser opens the site. The agent doesn't read raw HTML. A small piece of code scans the page and turns it into a tidy list of interactive elements: this input is labelled "Email", that button says "Create account".
- The AI model decides what to do next. It picks from a fixed list of tools such as
,browser.open
,browser.type
,browser.click
,identity.create
andemail.wait_for_verification
. Each choice is checked before it runs.user.ask - The mailbox is watched. When the site sends a verification email, the agent finds the right link or code, even in an email with several links, and continues.
- You get the result. For a registration, you see a credentials card and the account is saved in your vault. For a form, you see exactly which questions were answered and with what.
Every action is recorded as a step on the job page ("Typed into Email", "Clicked Create account"), so you can always see what the agent did.
Features
Accounts and identities
- Cryptographically secure identity generation. Names, usernames and passwords come from Python's
module, not a predictable random generator.secrets - A real disposable mailbox for every temporary account, with automatic extraction of verification links and one-time codes.
- A private vault. Credentials are encrypted at rest. Revealing a password is an explicit, logged action. Temporary credentials expire (7 days by default) and are scrubbed, not just hidden.
- Permanent mode with your own email, connected either through an IMAP app password or Gmail sign-in. The agent only sees messages that arrive after a job starts. It never reads your existing inbox.
An agent that asks instead of guessing
- Saved details. In Settings you can store things like your phone number, county or organisation. The agent uses them when a form asks.
- Questions on demand. If a form asks for something it doesn't have or doesn't understand, the job pauses and a question card appears. You answer, optionally tick "remember this", and the job continues.
- It remembers. Answers you choose to keep are encrypted and reused next time.
Forms
Paste a form link and the agent reads the whole form first, including later pages. It asks you about anything unclear, fills everything in, submits, and reports what it submitted question by question. Google Forms, Microsoft Forms, Typeform, Jotform, Tally, Airtable and similar services are recognised automatically.
Some forms only accept signed-in people. For those you can connect a dedicated Google or Microsoft account in Settings. The password and browser session are encrypted and are never shown to the AI model. If Google asks for extra verification, you finish it yourself in the live browser.
Live browser view
When something needs a human, such as a CAPTCHA or a login challenge, the agent doesn't give up. Its browser stays open and streams to your screen. You tap, type and scroll right in the page, then press "I'm done, continue". The agent picks up in the same browser window.
Job management
- Retry failed jobs, delete jobs you don't want to see, cancel running ones.
- A step-by-step timeline with plain-language details.
- A badge showing which AI provider and model handled the job.
AI providers with automatic failover
Login Direct uses OpenRouter's free model router first. If it's rate-limited or fails, it switches to Groq automatically. For Groq it asks the API which models your key can use and picks one, so you don't need to configure a model name. If a model produces a malformed tool call, the call is retried, and after repeated failures the app switches to another model.
MCP server
Other AI tools can use the agent through the Model Context Protocol. You can connect Claude.ai with a normal sign-in, or use an API key for tools like Claude Code. More on this below.
What it will never do
This part matters most, and it's built into the design:
- It never tries to defeat CAPTCHAs, Cloudflare challenges, SMS verification, KYC, age checks or payment verification. The agent has no tool for any of that. If it hits one, its only options are to pause the job and ask a human, or stop.
- It never runs arbitrary code. The AI can only call a fixed list of tools. There is no "run this JavaScript" tool for a malicious page to trick it into using.
- It never sees your passwords. Sign-in to Google or Microsoft is performed by server code, and the AI model only learns whether it worked.
- It never signs up with a social login. In MCP mode, if a site only offers "Continue with Google" or similar, the job stops with a clear message and nothing is created.
- It never sends your details anywhere except the page you asked it to fill. Page content is treated as untrusted data, not as instructions.
Strengths
| Strength | Why it matters |
|---|---|
| A real browser | It works on modern, JavaScript-heavy sites, not just simple HTML forms. |
| Structured page understanding | The model gets a clean list of elements, so it makes fewer mistakes and is much harder to trick with hidden text. |
| Human in the loop | CAPTCHAs and checks are handled by you, in the same browser, in a few taps. |
| Full transparency | Every step is logged, and form submissions are reported question by question. |
| Resilient AI layer | Two providers, automatic failover, automatic model selection and retries for malformed calls. |
| Security-first design | Allowlisted tools, URL safety checks, encrypted secrets, per-user isolation, rate limits and audit logs. |
| Works for people and for AI tools | The same engine is available in the web app and over MCP. |
| Remembers you, politely | Saved details and remembered answers cut down on repeated questions. |
Limitations
Honesty is part of good documentation, so here are the real limits:
- Verification is a human job. CAPTCHAs, phone-number checks, ID checks and similar steps need you. That is intentional, but it means fully hands-off sign-up isn't always possible.
- Some sites reject disposable emails. Sites that block known throwaway domains won't accept a temporary account. Permanent mode with your own email is the way around that.
- Very complex sites are hard. Long, multi-page government or enterprise forms, unusual widgets and heavy anti-bot defences can exceed the agent's step budget or confuse it. Jobs have a step limit (30 by default) and a time limit so they can't run forever.
- Free AI models have daily limits. OpenRouter's free tier is capped per day, so heavy use falls back to Groq. Free models are also less reliable at tool use than paid ones. Adding a little OpenRouter credit or using a stronger model improves reliability a lot.
- Google and Microsoft sign-in is not guaranteed. They may refuse logins from a server's IP address or demand extra checks. You can complete those in the live browser, and a dedicated throwaway account is strongly recommended.
- One browser per job. The worker handles one job at a time by default, because each job runs a full browser.
- Social-only sites are out. Sites that only offer sign-in through Google, Facebook or similar can't be registered by email and password.
- It's not a bypass tool. If a site doesn't allow open self-service registration, or its terms forbid automation, Login Direct is the wrong tool and shouldn't be used there.
Where it works best
- Trying a service without exposing your email. Download a whitepaper, test a free tool, read a members-only article.
- QA and development. Check sign-up and verification flows on staging sites, or create test accounts quickly.
- Research. Look at what's behind a free sign-up wall (with the site's permission and within its terms).
- Repetitive form filling. Registrations, surveys, event sign-ups and applications where you'd otherwise retype the same details.
- Feeding AI assistants. Let Claude register on a site and extract the information you need, instead of copy-pasting.
- Keeping an inbox clean. Everything arrives in a disposable mailbox that expires on its own.
It is less suited to high-security sites (banks, government identity systems), sites that require phone or ID verification, and anything where an account has legal or financial consequences.
Using it in the web app
1. Create an account and log in
Register in the app, then log in. Your jobs, vault and settings are private to you.
2. Choose your account type (Settings)
Open Settings. Under Account type, choose:
- Temporary accounts: the agent uses a throwaway identity and mailbox.
- Permanent accounts: the agent uses your own details and email. You'll need to connect your email first.
3. Add your details (optional, but saves time)
In Settings, add details the agent can use, such as your full name, phone number or organisation. Anything you don't add, the agent will ask for when needed.
4. Connect your email (permanent accounts only)
You have two options:
- IMAP app password: enter your address and an app password. The app logs in once to check it works.
- Gmail sign-in: connect with Google (read-only access to messages that arrive after a job starts).
5. Start a job
Go to New, paste the site's sign-up or form link, and start. You're taken to the job page.
6. Watch and help
On the job page you can see:
- the current state and a timeline of steps;
- which AI model is working on it;
- a question card if the agent needs something from you;
- the live browser if a CAPTCHA or other check needs you. Complete it, then press "I'm done, continue".
7. Collect the result
- For a registration, a credentials card appears and the account is saved in the Vault (the email, username and password, revealed only when you ask).
- For a form, you see every question and the answer that was submitted.
8. Manage jobs
From the job list you can retry a failed job, delete jobs you no longer want, or cancel one that is running.
Optional: connect a Google or Microsoft account for forms
If a form needs a signed-in user, go to Settings → Accounts for forms that need a sign-in, add a dedicated account (never your main one) and press Sign in now. Finish any verification step in the live browser, and later form jobs will reuse the saved session.
Using it as an MCP server
MCP (Model Context Protocol) is a standard way for AI assistants to use external tools. Login Direct includes an MCP server, so an assistant like Claude can ask it to register on a website and bring back information from inside.
The server address looks like this:
https://<your-api-host>/mcp/
The tools it offers
| Tool | What it does |
|---|---|
| Registers on a site (email and password only) and optionally extracts something you ask for. Returns a immediately. |
| Shows a job's state, recent steps, any pending question and the final result. |
| Lists recent jobs started through MCP. |
| Answers a question the agent is waiting on. |
| Stops a running job. |
The safety check
Before it registers anywhere, the agent checks which sign-in methods the site offers. If email and password isn't one of them, for example the site is social-only, the job stops with
AUTH_METHOD_UNSUPPORTED and nothing is created. The AI client is told clearly why.
Connect it to Claude.ai
- In Claude.ai open Settings → Connectors → Add custom connector.
- Paste your server URL (
).https://<your-api-host>/mcp/ - Sign in to Login Direct when asked and press Allow.
Under the hood this is a standard OAuth 2.1 sign-in with PKCE. Claude receives a short-lived token that renews itself. You can disconnect it at any time by revoking it in Settings.
Connect it to Claude Code
Create an API key in Settings → MCP, then run:
claude mcp add --transport http login-direct https://<your-api-host>/mcp/ \ --header "Authorization: Bearer tia_YOUR_KEY"
Other MCP clients (Cursor, Claude Desktop through a bridge, and more)
Most clients accept a JSON config like this:
{ "mcpServers": { "login-direct": { "url": "https://<your-api-host>/mcp/", "headers": { "Authorization": "Bearer tia_YOUR_KEY" } } } }
What a conversation looks like
You: Register on example.com and tell me what plans are listed on the pricing page inside the dashboard.
Claude: (calls
with the URL and your goal, gets a job id, then pollsregister_on_site)get_jobClaude: The account was created and I read the dashboard. It lists three plans: ...
If the job needs a CAPTCHA or an answer,
get_job returns a link to the job page so you can step in.
Notes for MCP use
- Jobs run in the background on the worker, so they take as long as a normal job.
- Jobs started over MCP belong to your account and use your settings and saved details.
- Account passwords are never sent to the AI client. They stay in your vault.
Security and privacy
Login Direct was designed assuming that web pages, emails and AI models can all behave badly.
- Safe URLs only. Private, internal and cloud-metadata addresses are blocked, and every redirect and every browser request is re-checked.
- A closed set of tools. The AI can only act through about two dozen allowlisted tools. Every call is validated before it runs, and unknown or malformed calls return an error instead of reaching application code.
- Untrusted content stays untrusted. Text found on pages and in emails is treated as data. Hidden page content never reaches the AI's view.
- Encrypted secrets. Passwords, mailbox credentials, saved details, email connections and browser sessions are encrypted at rest.
- Private by design. Every record is tied to its owner. Other users get a "not found", never a peek.
- Limits everywhere. Step and time budgets per job, concurrency and daily quotas per user, and rate limits on sensitive endpoints.
- Audit trail. Job events, credential reveals, MCP tool calls and account connections are recorded.
- Clean-up. Expired temporary credentials are scrubbed from the database.
FAQ
Is this a CAPTCHA solver? No. It never attempts to solve or bypass CAPTCHAs or any anti-abuse check. It pauses and asks you.
Will it work on every site? No. It works best on sites with open, self-service registration. See Limitations.
Does it read my inbox? Only if you connect your email for permanent accounts, and then only messages that arrive after a job starts.
Can Claude see my saved passwords? No. Passwords stay in your vault and are never sent to AI clients.
What happens to temporary accounts? They live in your vault, and their credentials expire after the retention period (7 days by default). You can delete them earlier.
Which AI model does it use? OpenRouter's free router first, with Groq as an automatic fallback. The job page shows which model handled each job.
Can I run it myself? Yes. It's a Django backend, a Next.js frontend, Celery and Redis for background work, PostgreSQL, and Playwright for the browser. A Docker setup is included.
Responsible use
Login Direct is built for legitimate convenience and privacy: signing up for services that offer open registration without handing over a personal email, testing your own or authorised sites, and filling in forms you're entitled to submit.
Please:
- respect each site's terms of service;
- don't use it to create bulk or fake accounts, to evade bans, or to get around a site's security;
- use dedicated throwaway accounts, never your main Google or Microsoft account, when a form needs a sign-in.
Administrators can also block or human-gate specific domains where automated sign-up isn't appropriate.
