Claude Skill Indexedai

Skill: /indexedai

Query any website via its MCP endpoint — structured data, zero context overhead. MCP loads only when invoked, never at startup.

Trigger

• /indexedai [question] — query a specific site
• /indexedai [question] — domain-only shorthand
• /indexedai — when a URL is implied by context

Flow obbligatorio

Before fetching anything, tell the user: > "Checking for MCP server at ``…"

Step 1 — Discover MCP via .well-known

GET https:///.well-known/mcp.json

Success — response contains:

{"mcpServers":{"name":{"url":"https://..."}}}

→ extract MCP endpoint URL, proceed to Step 2.

Failure (404, timeout >3s, malformed JSON, network error) → skip to Fallback.

Step 2 — Initialize (mandatory handshake)

POST 
Content-Type: application/json

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","clientInfo":{"name":"indexedai-skill","version":"1.0"}}}

If response missing result.capabilities or status ≥ 400 → skip to Fallback.

Step 3 — List available tools

POST 
Content-Type: application/json

{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}

Read result.tools[]. Pick the tool whose name/description best matches the user's question. If no tools returned → skip to Fallback.

Step 4 — Call the best tool

POST 
Content-Type: application/json

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"","arguments":{...}}}

Map user's question to tool arguments. Return result.content.

Fallback

Tell the user: > "No MCP server found at ``, using WebFetch."

Then use WebFetch on the most relevant URL for the question.

Multiple domains

If query contains multiple domains, run Step 1 for all domains in parallel before calling any tool. Then call each MCP concurrently.

Output format

Always prefix answers with source:

• [MCP: domain.com] — data from site's own MCP server
• [Web: domain.com] — fallback WebFetch

Example:

[MCP: docs.stripe.com]
Webhooks retry up to 3 days with exponential backoff...

Implementation notes

• HTTP stateless — no persistent connection to maintain
• No SDK or library needed — plain JSON-RPC over HTTP POST
• If initialize succeeds but tools/list fails → Fallback
• Never invent tool names — only use names returned by tools/list
• Respect result.content[].type: text = display as-is, resource = fetch if needed

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: guidodr27
• Source: guidodr27/claude-skill-indexedai
• License: MIT

Install and usage instructions live in the source repository linked above.