Login Machine

Why it's ranked here

The idea is sharper than most browser agents: one loop replaces a pile of per-site login scripts, and the README says it did exactly that inside Anon. The code backs the design claims. Every element address the model proposes is checked against the live page, including inside frames, before anything is clicked, and failures are fed back for up to three retries. Passwords go from the user to the page without ever being placed in a model prompt. It is MIT licensed and small enough to read in one sitting.

What's good

The markup trimming is careful: it drops scripts, styles, images and hidden elements, keeps only a short list of attributes useful for finding fields, and walks into shadow DOM and slots so embedded enterprise sign-in widgets still show up. Frame-aware fill and click helpers try the main page first, then every child frame. The model output is forced through a strict schema, and the instruction text tells it how to spot one-time-code fields and which locator styles to prefer. User actions stream progress back to the browser as server-sent events, so the form can show submitting and submitted states live.

Tradeoffs

It is a demo app, not a library: the package is marked private, there is no CLI or importable API, and no test script. It hard-requires paid BrowserBase and Anthropic accounts, with the model version pinned in code. After three failed validations it makes a fourth call and returns the result unchecked. The chat endpoint has no authentication, reflects any request origin in its CORS header, and accepts any non-empty string as the start URL. The instructions tell the model to skip Google, Apple and Microsoft sign-in buttons, so social login is out of scope. Cross-origin frames cannot be read.

How to use it well

Treat it as a reference design for your own agent's sign-in step rather than a dependency. Read the six step types and the validate-then-retry loop, then port them into whatever runs your automation. If you deploy the app as is, put it behind your own authentication and restrict origins first, and note that product analytics record the target domain whenever the analytics key is set. It fits flows built on email, password, one-time codes, account pickers and magic links; it does not cover federated social sign-in or unattended runs, since a person answers every prompt.

Observed

License

MIT (LICENSE file)

Language

TypeScript, Next.js app

Packaging

package.json marked private; scripts are dev, build, start, lint only

Interfaces

Web UI plus one POST endpoint with start, submit and close actions

Required services

Anthropic API key and BrowserBase API key plus project ID (.env.example)

Optional telemetry

PostHog key and host in .env.example; hook captures target domain

Tests

No test script or test framework in package.json

Validation retries

Up to 3 locator-validation attempts, then an unvalidated fallback call (src/lib/ai-login/agent.ts)

HTML cap

Extracted page markup truncated to 100,000 characters (src/lib/ai-login/browser.ts)

Excluded flows

Prompt excludes Google, Apple and Microsoft social login buttons (src/lib/ai-login/prompts.ts)

Endpoint access

No authentication check; CORS header echoes request origin (src/app/api/chat/route.ts)