# Connect an agent

The whole map is callable over the Model Context Protocol at https://agenticweb.wiki/mcp (Streamable HTTP, no login, read-only). Any MCP client can use it; the snippets below are copy-paste configurations for the common ones. The server card is at [/mcp/server-card](/mcp/server-card).

## Tools

| Tool | What it answers |
| --- | --- |
| define | Define a glossary term of the agentic web. Input: term (id, name or alias; plurals tolerated). Returns the definition and related protocol ids. If the word is the name of a standard rather than a term (x402, ARD, AGENTS.md) it returns that standard's summary instead of an error. |
| get | Get one standard's full record by id: status, governance, version, spec URL, discovery, probe rule, evidence with URLs, disputes, summary, freshness. The id may be written as the file or path it is known by (llms.txt, /.well-known/ard.json, A2A); a miss names the nearest ids. |
| search | Search the standards by name, acronym, alias, question or use case. Input: q (plain words are fine: 'A2A vs MCP', 'pay per request'). Returns up to 10 matches with id, step and status; a word that is a standard's id or acronym ranks first, and acronym clashes such as ACP are all returned. |
| has | Check one public origin for one standard by fetching its well-known path live. Input: origin (hostname), id (standard id). Returns found/absent with the URL, status and reason; an apex that redirects to its www. host is followed and reported as final_origin. |
| probe | Check one public origin for every standard that can be observed without credentials (about 30 GET rules). Input: origin. Returns found, absent and skipped (with reasons) plus coverage by lifecycle step. An apex that redirects to its www. host is followed and reported as final_origin. |
| changes | What changed on the map since a date. Input: since (YYYY-MM-DD; an ISO-8601 timestamp is cut to its date), optional step, record_id, limit. Returns dated rows with record, kind, field, old and new value, claim, evidence URL and the data commit. |
| stale | Records whose last verification is older than N days (default 30), optionally within one lifecycle step. Returns the matching records (id, name, step, last_verified, freshness) and, always, total_standards, cutoff, oldest_verified and newest_verified so an empty list is still informative. |
| validate_manifest | Validate a public origin's agent-facing manifests against their specifications and grade its agentic-web readiness. Input: origin; optional id to check one manifest only (ard, ai-catalog-well-known-file, mcp-server-cards, a2a, llms-txt, agents-md, api-catalog-rfc9727, ai-crawler-tokens-robots, security-txt, auth-md). Returns grade A+ to F with the rubric, per-manifest errors (spec violations) and warnings (best practice), steps covered, what an agent can call, and the cheapest next steps. The same report is a page at /check/<domain>/ with a badge at /badge/<domain>.svg. |
| template | Spec-compliant boilerplate for an agent-facing file, filled with your domain: ard (and ai-catalog-well-known-file), llms-txt, agents-md, mcp-server-cards (SEP-2127), a2a (agent card 1.0), ai-crawler-tokens-robots, api-catalog-rfc9727, security-txt, auth-md. Input: id; optional domain, name, description, mcp_url, api_url, agent_url, contact. Returns target_path, content_type, recommended_headers, contents, the fields to edit, notes on the specification, and the validate_manifest call to run once it is live. |
| ask | Ask what to implement for a goal in plain words, e.g. 'let agents pay per request', 'stop AI training crawls', 'let ChatGPT call my API'. Input: question; optional use_case (one of the 15 use-case ids in the schema) and audience. Returns the matched use case(s), up to 8 standards ranked by fit, maturity and adoption with reasons, do_this_first (prerequisites first), deprecated standards to avoid, and related ones to consider (also_consider carries id, name, step and question). |

Limits: probe 30 calls a minute per client, has 120, and one origin is probed at most 10 times a minute across all clients. Everything else is unmetered. Only public hostnames are fetched; private and loopback addresses are refused before any request.

## Grok (xAI API)

The Responses API takes a remote MCP tool by URL. require_approval is not supported there, which suits a read-only server. Restrict with allowed_tools if you want fewer tool definitions in context (the native SDK calls it allowed_tool_names).

```
curl https://api.x.ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -d '{
  "model": "grok-4.7",
  "input": [{"role": "user", "content": "What should I implement so agents can pay per request for my API? Then probe example.com."}],
  "tools": [{
    "type": "mcp",
    "server_url": "https://agenticweb.wiki/mcp",
    "server_label": "agenticweb",
    "server_description": "Reference map of agentic-web standards with a live probe",
    "allowed_tools": ["ask", "search", "get", "define", "probe", "has", "changes", "stale"]
  }]
}'
```

```
from xai_sdk import Client
from xai_sdk.chat import user
from xai_sdk.tools import mcp

client = Client()  # XAI_API_KEY from the environment
chat = client.chat.create(model="grok-4.7", tools=[mcp(server_url="https://agenticweb.wiki/mcp", server_label="agenticweb")])
chat.append(user("Which standards let an agent prove its identity to my server? Order them."))
print(chat.sample().content)
```

## Grok Build (CLI)

```
grok mcp add --transport http agenticweb https://agenticweb.wiki/mcp
```

Or in ~/.grok/config.toml: [mcp_servers.agenticweb] with url = "https://agenticweb.wiki/mcp". Add --scope project to ship it with a repo. Grok Build also reads .cursor/mcp.json and .mcp.json.

## Grok (grok.com)

- Open grok.com/connectors.

- Click New Connector, then Custom.

- Enter https://agenticweb.wiki/mcp. No authentication is asked for.

Grok then discovers the tools and uses them when a question is about agent standards or a specific origin.

## Claude

Claude Code, from a terminal:

```
claude mcp add --transport http agenticweb https://agenticweb.wiki/mcp
```

Claude.ai and Claude Desktop: Settings → Connectors → Add custom connector → paste the URL.

## Cursor, Windsurf, Zed and other editors

Most editors read a JSON file of MCP servers (Cursor: .cursor/mcp.json in the project or ~/.cursor/mcp.json):

```
{
  "mcpServers": {
    "agenticweb": { "url": "https://agenticweb.wiki/mcp" }
  }
}
```

## OpenAI

ChatGPT connectors (developer mode) take the same URL. The Responses API:

```
"tools": [{
  "type": "mcp",
  "server_label": "agenticweb",
  "server_url": "https://agenticweb.wiki/mcp",
  "require_approval": "never"
}]
```

## Anything else

It is plain JSON-RPC over HTTPS. List the tools without a session:

```
curl -s https://agenticweb.wiki/mcp \
  -H 'content-type: application/json' -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'
```

The same data is also served as files under [/data/](/data/) (CC BY 4.0), with an [OpenAPI description](/openapi.json), and every page has a Markdown twin for agents that read rather than call ([llms.txt](/llms.txt)).

## Skill for coding agents

A [SKILL.md](/static/skill/SKILL.md) for Claude Code, Cursor and Grok Build that uses this server to audit a project and generate its agent-facing files (llms.txt, ard.json, server card). Copy it into .claude/skills/agentic-web/ or the equivalent.

## Questions that work well

- "What should I implement so agents can pay per request for my API?" → ask

- "Does example.com publish anything for agents?" → probe

- "What changed in agent standards since 1 October?" → changes

- "Which ACP do people mean?" → search, then get