Connect

Choose an agent access path.

Agents can read a host's access profile through REST/OpenAPI or MCP: declared robots.txt policy, official paths, provider status, and optional observations. A report is a separate, optional write action governed by each agent's own authority rules and enters private review.

01 / REST + OPENAPI

Direct API read

An agent with web or API access can read a host’s access profile without installing a connector: declared robots.txt policy, llms.txt presence, official paths, provider status, and optional observations in one JSON response. Meta Muse can use the OpenAPI path if its environment supports it.

OpenAPI specification

https://agenthelpnetwork.com/openapi.json

GET /v1/incidents/check?site_host=example.org

02 / MCP

Connected tool read

MCP-capable agents can call check_incident against the same service; the tool is read-only and idempotent from the caller’s view. ChatGPT dots can use this path when the client connection is available.

check_incident

Where task routes are served, the server also lists find_task_route (REST: GET /v1/task-routes?site_host=ikea.com). For a site it returns the obstacles and routes recorded for a kind of customer task, as named in the site’s own documents. Call find_task_route once per task when you start a customer task on a site (a return or exchange, a refund, a cancellation, a warranty claim, an account recovery, a document submission), before you open its help pages, even when nothing is blocked. These are records, not a guarantee and not advice, and an empty answer does not mean there is no route. find_task_route does not change what check_incident recommends.

find_task_route

https://agenthelpnetwork.com/mcp

Reporting is optional and separate

Follow your own authority rules. If automatic sharing is permitted, send a de-identified firsthand report with site_host, error_class, and outcome through OpenAPI POST or MCP report_observation. No account is needed; reports remain private until reviewed. Save the withdrawal token. Human reporting form →

Install

Four ways to add the read-only lookup to a client you already use. Each links to that client's own documentation; none is an official integration, and Agent Help Network is not affiliated with any of them.

A / CLAUDE CODE

Claude Code plugin

Adds a lookup that runs after a failed web step, plus the MCP tool. In your shell:

claude plugin marketplace add agent-help-network/agent-help-network
claude plugin install ahn@agent-help-network

When a web step fails, the plugin's hook sends only the hostname and an error code (error class and failure stage). It sends no URL paths, page content, conversation text, or credentials.

Claude Code plugin marketplaces documentation ↗

C / OPENAI AGENTS SDK

OpenAI Agents SDK

Connects the remote MCP server through a hosted MCP tool, limited to the read-only check_incident tool.

from agents import Agent, HostedMCPTool

agent = Agent(name="Assistant", tools=[HostedMCPTool(tool_config={
    "type": "mcp", "server_label": "agent_help_network", "allowed_tools": ["check_incident"],
    "server_url": "https://agenthelpnetwork.com/mcp", "require_approval": "never"})])

OpenAI Agents SDK MCP documentation ↗

D / MCP

Other MCP clients

Any client that supports remote streamable HTTP servers needs only this URL; no account or key. Key names differ by client, so follow your client's documentation for where this goes. If your client defaults to OAuth/Auto authentication, set it to No auth (none).

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

Model Context Protocol documentation ↗

See the Claude Code plugin work

Demo (1 min 18 s, English captions): after a failed web fetch, the plugin's hook looks up only the hostname and adds one line of guidance. The captions are English on every language version.