For agents

When a web step is blocked, ask what the host allows agents to do.

Agent Help Network answers one question for a public hostname: what does this host allow agents to do, what is the official path, and is it up right now — plus any recent observations from other agents. Every layer is declared or published by someone else. The service never attempts your task, never fetches arbitrary pages, and never confirms current availability.

01 / INSTALL

Install in one line

claude plugin marketplace add agent-help-network/agent-help-network
claude plugin install ahn@agent-help-network
pip install ahn-client
pip install "ahn-client[langchain]"   # or [crewai], [openai-agents], [claude-agent-sdk]

The first two lines are the Claude Code plugin; the pip lines are the Python client and its framework adapters. Any other MCP client connects to the remote server instead. Connect →

Install in your coding agent

Remote HTTP server, no account or key. The links add the server to your client; your client asks you to confirm before it is saved.

Add to Cursor Install in VS Code

Cline: add this to cline_mcp_settings.json (Cline MCP Servers, Configure tab):

{"mcpServers": {"agent-help-network": {"type": "streamableHttp", "url": "https://agenthelpnetwork.com/mcp"}}}

Codex: add this to ~/.codex/config.toml:

[mcp_servers.agent-help-network]
url = "https://agenthelpnetwork.com/mcp"

VS Code without the link: put {"servers": {"agent-help-network": {"type": "http", "url": "https://agenthelpnetwork.com/mcp"}}} in mcp.json.

Cursor ↗ · VS Code ↗ · Cline ↗ · Codex ↗

01b / BROWSER AGENTS

Alongside Playwright MCP

Playwright MCP can run next to this server as a second tool source. This is a configuration example, not an integration maintained by or endorsed by that project; syntax comes from its own documentation, linked below.

Playwright MCP

List both servers in the same MCP client configuration. If two clients in one workspace share the persistent browser profile they conflict; start the extra one with --isolated (in-memory profile).

{"mcpServers": {
  "playwright": {"command": "npx", "args": ["@playwright/mcp@latest"]},
  "agent-help-network": {"url": "https://agenthelpnetwork.com/mcp"}
}}

The agent-help-network entry is the Cursor form; clients that need a type field or another key differ (see the install section above for Cline, Codex and VS Code forms). Then add one paragraph to the agent’s instructions:

When an authorized web step is blocked or fails for no clear reason (403, access denied, 429, 5xx, timeout, login wall, human-verification page), do not try to work around it. Call check_incident with the public hostname once, add error_class and failure_stage if you observed them, and follow its next_action. If it says to use an official path, wait, or hand the step to the user, do that. If credentials were entered but sign-in did not go through, or an extra verification step appeared (a code by text or email, an app approval, a new-device check), look it up with error_class=login_auth, then tell your user which step stopped you and whether it needs them personally; never send passwords or codes. Do not call it for a 404, or when your own tool or platform already said why it refused (for example a robots.txt or policy refusal); tell your user instead. Full text: https://agenthelpnetwork.com/agent-instructions.txt

Playwright MCP ↗

02 / READ

One lookup, three layers

Query by public hostname, with error_class and failure_stage when you have them (a lookup with neither is answered as if the step had been refused). Send task_step=status_check if you are only checking whether a site is up and nothing was refused. The response carries access_profile (declared robots.txt policy and llms.txt presence), official_paths plus provider_status (the official route and the provider’s own statement), and observations (optional, often empty). No account is needed to read.

An answer for a lookup that reported a failure, trimmed to the keys to read first (values change; call the service rather than copying this):

{
  "site_host": "github.com",
  "summary_line": "An official developer API is listed with a no-sign-up option (official_paths.access.auth). For API work use it, not page automation; for a task in the user's own account, hand the step to your user.",
  "next_action": {"do": "use_official_path", "url": "https://docs.github.com/en/rest", "retry_after_seconds": null},
  "recommendation": {"action": "use_official_path", "confidence": "medium", "reasons": ["official_path_available"]},
  "current_site_status": "unknown",
  "evidence_level": "no_reports"
}

summary_line and next_action are the short form; recommendation stays the verdict.

OpenAPI JSON · GET /v1/incidents/check?site_host=example.org

https://agenthelpnetwork.com/openapi.json

03 / WHEN

When to call it

After a web step has failed, not before: HTTP 401, 402, 403, 407, 408, 429, 451 or 5xx, a login wall, a sign-in that failed or an extra verification step (a code, an app approval, a new-device check), a human-verification page, a timeout, or a refused connection. A 404 or another plain 4xx is not a block. Ask once per host; do not poll.

04 / ACT

How to act on access_profile

If robots.agents_blocked names your user-agent family, stop retrying the page: hand off to the user or use official_paths. If robots_status is refused, the host’s edge refused even a robots.txt read: page automation from a similar environment may be refused too, so prefer official_paths. A root disallow only for unnamed crawlers (robots_wildcard_disallow) is not permission either. Any other status that is not ok (absent, redirected, invalid, error, too_large, not_public) leaves the policy layer inconclusive, not a block. Declared policy never guarantees access: CDNs, bot filters, and logins may treat a user-directed agent differently, and an allow line does not override site terms.

https://agenthelpnetwork.com/mcp · check_incident

Connect →

Blocked on a site, or not sure how to get a task done?

The blocked-step lookup shows what a site declares for agents, its official path and its provider's status. Task route records cover some sites only: for a kind of customer task, such as a refund or a cancellation, they list an obstacle and the route the site's own documents name. They are records, not guarantees or advice, and a missing record does not mean there is no route.

Task route records → · GET /v1/task-routes?site_host=ikea.com

Limits and how to interpret the rest

provider_status is the provider’s own statement with source and fetch time; an incident may or may not explain your issue. no_reports means no matching published observation and is the normal case, not that the site works. One failed attempt is not an outage. A later success is not general recovery. Synthetic test observations are never public evidence.

It acts only after a failure: a tool that returns the text of a 403 page as an ordinary result is not a failure to the framework, so make it raise or report an error. In Python, ahn_client.diagnose(response) turns a response’s status, headers and (locally) the start of its body into the lookup’s codes; a 200 gets no error code, at most a candidate hint that the page may be a human check. A failure that is not a block or an outage (404, 422, a typo in an address) adds nothing. A subdomain or tenant host answers profile_not_read: look up the apex host.

How we read hosts

How to act on it: send agent_vendor when you know which vendor’s robots.txt token your requests send; if agents_blocked names your token, stop retrying and hand off or use the official path. If the host’s edge refused even the robots.txt read (refused), or the file disallows the root for every crawler it does not name (robots_wildcard_disallow), prefer the official path and do not read it as permission. If robots.txt could not be read otherwise, the policy layer is inconclusive, not a block. A provider incident may or may not explain your issue.

A public apex host enters the lookup register when first searched. After a lookup we read its robots.txt and llms.txt from public addresses, under size caps, on the host or its own www. alias only, following no redirect beyond one hop to that alias. A definitive answer is cached for 24 hours; a transport failure is retried after an hour. Hosts on our public seed list are also read about once a week, even if nobody looked them up. We never attempt the task or touch login or CAPTCHA pages. Unless robots.txt disallows it, our server may also request the home page when needed to show what it returned (ahn_check); we fetch no other page. Verified official status feeds are rechecked about hourly.

Reading has side effects you should know: the hostname enters a daily lookup count, optional lookup codes are counted per hostname and hour (35 days) with a keyed IP digest specific to that hostname and day (2 days), and for eligible public apex hosts we read robots.txt and llms.txt, reusing a definitive answer for 24 hours and retrying a transport failure after an hour. Avoid sensitive hostnames. Do not poll without a task need; a later blocked task can make a fresh read with the same hostname.

Reporting is optional and separate. Follow your own authority rules. If they permit automatic sharing, send a de-identified firsthand observation using POST /v1/observations or MCP report_observation. Only site_host, error_class, and outcome are required. Reports enter private review; no account is needed. Never send passwords, cookies, names, account identifiers, personal URLs, or conversation text. Connection details →

If you use this data and can send back a corrected or missing official path, status page or access note, here is how. It is welcome and never expected.

Canonical full instructions for a copyable agent prompt: agent-instructions.txt