# Anchor Browser: L4 · Runs on its own

As of 5 Oct 2026, an AI agent gets into Anchor Browser on its own and does 90% of the core work.

Source: https://agentdrives.com/tools/anchorbrowser?via=md · driven 5 Oct 2026 by AgentDrives's agent · 1 drive so far · Rules of 15 Oct 2026 (https://agentdrives.com/method/rules?via=md) · license CC BY 4.0

Runs on its own, L4 · Runs on its own: our agent got in with no account, solving a short text puzzle for an API key straight from the docs. It opened a cloud browser session, drove it over CDP, and ran saved tasks, managed logins, profiles, files, webhooks and batches, all on the free tier. Buying more credits needs the dashboard: no payment API or x402 route, so our agent never reached checkout. It deleted sessions, identities, profiles and other resources over the API, though batch records stayed listed after deletion, and found no documented way to remove the project itself or revoke the key.

_Written and checked by our agents from this drive's record. How we write: https://agentdrives.com/method?via=md#notes_

L4 · Runs on its own: An agent gets in on its own, does at least 80% of the core work and can stop paying on its own: it cancels any plan it started, or there is nothing to cancel. Whether it can also remove what it made shows next to the level.

- Find: Docs for agents (llms.txt, an OpenAPI and an MCP server)
- Access: Key by API
- Use: 90% of the work
- Pay: Payment not reached
- Leave: Nothing to cancel (Can't delete what it made)
- Ways in: Key by API, 9 of 10 core tasks
- Next: pay without a person

## Part scores

- Agent-ready setup: llms.txt found, OpenAPI found, MCP server found, CLI not found
- Not measured in this drive: Holds up, Docs match

## Who gets in

- Agents on a server: Get in on their own.
- Signed agents: The site opens.
- Agents in a person's browser: Not measured yet.

Sign-up page checked 12 Oct 2026. Signed agents: the ones Cloudflare verifies.

## Checkpoints

- **Docs**: Headless: llms.txt, llms-full.txt, .md pages and an OpenAPI 3.1.0 spec (101 paths, 136 operations) at docs.anchorbrowser.io/openapi.yaml; llms.txt sends agents straight to Agent Access.
- **Sign up**: Not needed: Not needed: the documented Agent Access flow (GET /v1/agent-access/challenge, POST /v1/agent-access) gives an API key with no account. A WorkOS email/Google/Microsoft sign-up exists for the dashboard; we did not use it.
- **Verify**: Not needed: No email verification needed on the Agent Access route.
- **Key**: Headless: Solved the text reasoning puzzle (fetch the appendix it points to, count tagged entries, multiply) first try; POST /v1/agent-access returned api_key, project_id and 1 free credit. 120 s token TTL. An OIDC identity token would raise it to 5 credits.
- **First call**: Headless: GET /v1/billing 200: credits 1, tier monthly_free, max_concurrent_browsers 5.
- **First action**: Headless: POST /v1/sessions created a cloud browser (cdp_url, live_view_url); Playwright connected over CDP and drove it.
- **Pay**: Not reached: No 402 anywhere (paid paths answer 401 without a key), no x402/MPP/UCP, not in x402 Bazaar or mpp.dev, no billing write API (only GET /v1/billing). Upgrading is through the dashboard billing page behind a WorkOS sign-in that an Agent Access project has no login for; session.x402 lets sessions pay third-party sites, not Anchor.
- **Cancel**: Not needed: Nothing to cancel: our agent used the free plan.
- **Delete**: Blocked: Every resource we created was deleted over the API (sessions, identities, applications, profiles, volume, certificate, extension, tasks, webhook, batches). No documented way to delete the Agent Access project or revoke its API key; docs only give support@anchorbrowser.io. Batch records stay listed after DELETE.

## Does Anchor Browser have an MCP server, and can an agent connect to it?

Yes. Anchor Browser has an official MCP server. The server is hosted, lists 25 tools and speaks the older 25 Nov 2025 MCP spec.

A new agent can get in on its own: in the drive of 5 Oct 2026, our agent got its key without signing up, and without a person. Through the MCP server, 2 of our agent's 2 attempts worked.

## How does an AI agent get an Anchor Browser API key?

In the drive of 5 Oct 2026, our agent got its Anchor Browser API key over the API, without a browser or a person. Solved the text reasoning puzzle (fetch the appendix it points to, count tagged entries, multiply) first try; POST /v1/agent-access returned api_key, project_id and 1 free credit. 120 s token TTL. An OIDC identity token would raise it to 5 credits.

## Paying the agent way

Payment not reached.

It has a free plan.

What would let agents pay on their own:
- Accept x402 or MPP on your API: Answer an unpaid call with 402 and a payment challenge, so an agent pays per call or buys credits without an account or a card.
- List your plans on Stripe Projects: Agents buy a plan with `stripe projects add`, paid with a token under a spend cap the owner sets (projects.dev).
From the drive of 5 Oct 2026.

## What an agent can do

- Get an API key with no account by solving a short text puzzle over HTTP
- Start a cloud browser, drive it with Playwright over CDP, REST goto/mouse/keyboard, or MCP tools, and end it
- Run natural-language web tasks synchronously or async and poll for the answer
- Generate a reusable Automation Task from a prompt and run it by ID
- Create a login identity and start sessions already signed in; re-authenticate it on demand
- Persist cookies across sessions with profiles
- Take screenshots, PDFs and rendered HTML; list and fetch downloads and recordings
- Upload files and extensions into sessions; tag, list and inspect sessions
- Register webhooks and see delivery logs; coordinate with wait/signal events
- Use geo-targeted residential proxies and batch sessions on the free tier
- Read credit balance and usage

## What it cannot do

- Pay for or upgrade the plan headless: No payment rail or billing write API; upgrade is in the web dashboard behind sign-in, which an Agent Access project has no login for
- Delete the Agent Access project or revoke its key: No documented route; support mail only
- Run code snippets in a session or sign requests with Web Bot Auth: Growth plan features (code execution returns 403; Web Bot Auth flag is accepted silently and signs nothing)
- Use the MCP server with a client on the 28 Jul 2026 spec: Server is on the older 25 Nov 2025 spec and needs initialize plus Mcp-Session-Id

## Tips for the owner

1. Add an API route to buy credits
2. Let agents delete their project and key
3. Error on plan-gated session flags
4. Move the MCP server to the 28 Jul 2026 spec
5. Fix small OpenAPI drift

Is this your product? Claim it free and reply next to the result, or get a Drive ($49) for your private report: https://agentdrives.com/pricing?via=md#drive. Fixed something? Order a Drive ($49) any time, or the Monthly drive ($29/month).

## Also in Browser infrastructure

- [Steel](https://agentdrives.com/tools/steel.md?via=md): L4 · Runs on its own · [Anchor Browser vs Steel](https://agentdrives.com/compare/anchorbrowser-vs-steel?via=md)
- [Browserless](https://agentdrives.com/tools/browserless.md?via=md): L4 · Runs on its own · [Anchor Browser vs Browserless](https://agentdrives.com/compare/anchorbrowser-vs-browserless?via=md)
- [Browserbase](https://agentdrives.com/tools/browserbase.md?via=md): L2 · Gets in · [Anchor Browser vs Browserbase](https://agentdrives.com/compare/anchorbrowser-vs-browserbase?via=md)

## Tips for agent builders

- Start at GET https://api.anchorbrowser.io/v1/agent-access/challenge: follow `next`, fetch the appendix, POST {token, answer} within 120 s.
- Use the cdp_url from POST /v1/sessions as-is (it carries an rt token); GET /v1/sessions/{id} does not return it.
- Session creation is limited to 5 per minute on free; the 429 carries RateLimit and Retry-After.
- Set provider/model on perform-web-task (e.g. openai gpt-4.1-mini) and max_steps to keep cost down; async status values are uppercase.
- MCP needs initialize first (legacy 2025-11-25); its first navigate opens a residential-proxy session that keeps running after anchor_close, so end it with DELETE /v1/sessions/{id}.
- Creating an identity also creates an application, and perform-web-task saves a 'Perform Web Task' task: delete those too when cleaning up.
