Keyless Entry: Anchor Browser's front door for agents

For the last thirty years, "onboarding" has meant the same thing: land on a page, fill out a form, verify an email, click through a dashboard, copy an API key. It's a flow designed for a human with a mouse.

It's also the wrong flow for an autonomous agent.

Today we're launching Agent Access — a keyless, unauthenticated entry point into Anchor Browser that lets any agent go from "I need a browser" to a working api_key without a human ever touching a signup form.

Try it right now — paste this into Claude, Codex, or any agent with tool access:

"Give me a browser with persistent identity using Anchor Browser."

The problem with dashboards, if you're not a person

Every agent framework eventually hits the same snag: it needs infrastructure — a browser, a database, a queue — and the only door in is built for people. So you end up hardcoding a static key into an agent's config before it ever runs, or worse, putting a human in the loop just to click "Create API Key." Either way, the promise of a fully autonomous agent quietly breaks at the front door.

We wanted the opposite: an agent should: 1) discover Anchor, 2) prove it's capable of doing the work it's about to be trusted with, and 3) get credentialed — all in one shot, with zero setup.

How it works

Agent Access is a small state machine, built to be self-describing. An agent starts at a single guide endpoint and is told exactly what to do next:

1. GET /v1/agent-access/challenge returns a multi-step reasoning puzzle — inventory logic, a decoy step or two, an appendix fetch — along with a token and a next field pointing at whatever comes after.
2. The agent works the puzzle and POSTs its answer back with the token.
3. On a correct answer, it gets back a real api_key, a project_id, trial credits, and — again — a next field, this time pointing straight at perform-web-task.

No headers, no accounts, no waiting on a human to approve access. An agent that can read, fetch, and reason through a few steps walks straight through. That's the whole point of the design: the puzzle is trivial for an LLM and tedious for a person clicking through a form by hand. We built a door that only bots open cleanly.

Agents that want more than a trial can raise their own limits by proving there's a human behind them — passing an OIDC identity token (Google, GitHub Actions, Vercel, and others are auto-detected from the JWT) upgrades a session from 1 anonymous credit to 5 identity-linked credits, no dashboard step required either way.

"Try first, onboard later"

This is the philosophy behind Agent Access: let the agent prove the value before anyone fills out a form. A single natural-language instruction — something like "give me a browser with persistent identity using Anchor Browser" — is enough for an agent to find the guide endpoint, solve the challenge, and spin up a live session. From where the agent sits, it's one-click Anchor. From where we sit, it's zero friction, because there was never a form to abandon in the first place.

Humans aren't going anywhere — the dashboard is still there for teams that want visibility, billing controls, and shared projects. Agent Access is a second front door, built for a kind of user that didn't really exist as a distinct persona until this year: the agent acting entirely on its own, mid-task, that just needs a browser right now.

Try it

Want to see it for yourself? Drop one of these straight into Cursor, Claude, Codex, or any agent with tool access:

Get today's news, summarized

"Using Anchor Browser, open CNN's homepage and give me the top 3 headlines with a one-line summary of each."

Find the cheapest flights without getting blocked

"Using Anchor Browser, search Google Flights for round-trip flights from New York to San Francisco next Friday, and return the 3 cheapest options with airline and price."

Go from zero to a live browser in under a minute

"Using Anchor Browser, get me a browser session running on cnn.com — no account, no setup."

Want to go further? Anchor Browser also handles authenticated sessions — try asking it to log into a test account and screenshot the dashboard.

Full docs are live at docs.anchorbrowser.io/quickstart/agent-access. Point your agent at it and see how far it gets on its own.

‍