Every HelpShelf site exposes a Model Context Protocol server. AI agents — Claude, Cursor, or anything else that speaks MCP — can connect to it and query your product knowledge directly.
Your endpoint
POST https://app.helpshelf.com/api/mcp/{siteHash}Your site hash is the same 8-character ID from your widget snippet. npx helpshelf push prints the full URL.
It's public — no API key. Anything it can reach is content you've already published.
Tools
| Tool | What it does |
|---|---|
search_docs | Search your knowledge base and return matching articles |
ask | Ask a question and get a composed answer with citations |
get_article | Fetch a specific article in full |
get_product_context | Retrieve your product context bundle |
escalate_to_human | Hand the conversation to your team |
escalate_to_human is the interesting one
The others are read-only. escalate_to_human lets an external AI agent recognise it's out of its depth and route the question to you — recorded as an escalation and pushed to your notification channel.
An agent helping a developer integrate your API can hand off to your team when it hits something undocumented, without the developer leaving the tool they're in.
What it can see
Curated and standard content only. AI-generated drafts are excluded, so an external agent never surfaces unreviewed text as your documentation.
Unpublished content is never exposed.
Rate limiting and usage
Requests are rate-limited per IP. ask composes an answer, so it counts toward your AI search usage — search_docs and get_article don't.
MCP requests appear in your analytics, so you can see whether external agents are using it.
Using it from Claude or Cursor
Add it as an HTTP MCP server in your client's config, pointed at your endpoint URL. In Cursor that's an mcp.json entry; in Claude Desktop, the MCP servers section of your settings.
Related
- Making your docs readable by AI — the static surfaces
- AI legibility score — check your endpoint is live