# Agent-friendly help

> Verified against the product on 2026-10-06

## Give agents access

If your Help Center is public, anyone can read its published pages without an account, key, or cookie, and its public MCP address answers from those pages.

- Public Help Centers allow crawlers and set the Content Signal to `search=yes, ai-input=yes`. Team-only Help Centers disallow crawlers.
- Team members can read a team-only Help Center in a browser signed in to Cactus or with a token from that Help Center’s sign-in.
- Draft pages and pages meant for one person aren’t reachable through the Help Center or its MCP address.

## Machine-readable files

These paths are relative to your Help Center’s origin. When enabled, its **Agents** page lists the MCP address, any available A2A address, browser tools, and machine-readable files.

| File or address | What it provides |
| --- | --- |
| `robots.txt` | Public Help Centers allow crawlers and state `search=yes, ai-input=yes`; team-only Help Centers disallow crawlers. |
| `/sitemap.xml` | Page addresses and, when known, when each page last changed. |
| `/llms.txt` | Published pages, one line each, with guidance on when to read them. |
| `/llms-full.txt` | Every page in one file. |
| `/<page>/md` | A Markdown copy of a page. |
| `/openapi.json` | The Help Center’s OpenAPI description. |
| `/.well-known/mcp/server-card.json` | The MCP address, its sign-in requirements, and its tools. |
| `/.well-known/agent-card.json` | The A2A address and its skill, on public Help Centers. |
| `/.well-known/agent-skills/index.json` | The skill index and a digest of the skill file. |
| `/.well-known/agent-skills/ask-help/SKILL.md` | Instructions for answering from published help over MCP or as text. |
| `/auth.md` and `/.well-known/oauth-authorization-server` | Agent sign-in guidance and OAuth metadata. |
| `/memory.git` | A read-only clone of a public Help Center’s published pages, topic and related-page links, and public What's New releases. |

To clone a public Help Center, use its site address in this command:

```sh
git clone https://<site>/memory.git
```

- The API catalog associates the Help Center with its OpenAPI description, MCP address, and MCP card; public sites also include the A2A card.
- The AI discovery catalog lists MCP and, on public Help Centers, A2A.

## Search and read

- From an MCP client, use `search_help` to find pages about a word or feature, then `read_page` to read a page in full.
- In a browser, `search_help` finds matching pages and `read_help_page` reads a page by its slug.
- While your Help Center is live, browser agents can also list topics and read releases.
- `ask_cactus` returns an answer with the pages it cites.

## Test browser tools

1. On your Help Center’s **Agents** page, check the registration status for this browser.
2. On your Help Center’s **Agents** page, select **Run** beside a sample tool. The result appears beneath it.

- detail: Docs, **Run** (1)

*On the Help Center's Agents page, select Run beside a sample tool.*

If your browser doesn’t expose WebMCP tools or can’t list its registered tools, **Run** executes the same tool on the page. WebMCP requires enabling a browser flag; the **Agents** page lists steps for Chrome and Edge.