Ask this site questions — wire the MCP endpoint into your AI client.

One command — claude mcp add

From any terminal:

claude mcp add --transport http solo-demos https://solo-demos-mcp.tom-ed6.workers.dev/mcp

Then in a Claude Code session, run /mcp to confirm solo-demos is connected. The tools search and get_doc become available immediately — no restart.

To remove it later: claude mcp remove solo-demos.

Settings → Connectors → Add custom connector

On claude.ai (Pro / Team / Enterprise), open Settings → Connectors → Add custom connector and paste:

Name:  Master the Mesh
URL:   https://solo-demos-mcp.tom-ed6.workers.dev/mcp

Leave authentication off — the endpoint is public. Save, then enable the connector in any chat via the Tools menu (bottom-left of the composer).

Edit ~/.cursor/mcp.json

Cursor reads MCP servers from a JSON file. Add this entry:

{
  "mcpServers": {
    "solo-demos": {
      "url": "https://solo-demos-mcp.tom-ed6.workers.dev/mcp"
    }
  }
}

Restart Cursor. The MCP indicator (bottom-right of the chat panel) should show solo-demos with two tools loaded.

Generic MCP — streamable-HTTP transport

Anything that speaks MCP over streamable-HTTP can connect. The endpoint is the URL above; no headers, no auth, no SSE fallback needed.

# Smoke-test from the command line (no client required):
curl -s -X POST https://solo-demos-mcp.tom-ed6.workers.dev/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

You should see search and get_doc in the response.

Reuse the existing search index — don't rebuild

The site already publishes search-index.json on every push (built by scripts/build-search-index.py from each lab's rendered index.html). The MCP server doesn't index anything itself — it just fetches that same file:

// mcp/src/index.ts — the loader in full
async function loadIndex(env: Env): Promise<Index> {
  const res = await fetch(env.INDEX_URL, {
    cf: { cacheTtl: 300, cacheEverything: true },
  });
  return (await res.json()) as Index;
}

cf.cacheTtl: 300 tells Cloudflare's edge to keep the JSON cached for 5 minutes. Push new content → it's searchable within 5 min, with no worker redeploy.

Cloudflare's deploy CLI

Wrangler is an npm-installable CLI that bundles, uploads and routes your Worker. The whole config is one file:

# mcp/wrangler.toml
name = "solo-demos-mcp"
main = "src/index.ts"
compatibility_date = "2025-05-01"

[vars]
INDEX_URL = "https://www.masterthemesh.com/solo/search-index.json"
SITE_BASE = "https://www.masterthemesh.com/solo/"

name becomes the first segment of the public URL. [vars] values are injected at runtime as env.INDEX_URL / env.SITE_BASE — handy for swapping the index source without touching code.

Three commands

cd mcp
npm install                              # one-time
./node_modules/.bin/wrangler login       # OAuth in browser — one click
./node_modules/.bin/wrangler deploy

On a fresh Cloudflare account, the first deploy also asks you to pick a workers.dev subdomain (becomes part of every worker URL on your account — e.g. tom-ed6.workers.dev). End-to-end from git clone to public endpoint: about three minutes.

How Cloudflare actually runs it

Workers don't run in containers or VMs. The code runs inside a V8 isolate — the same sandbox Chrome uses for browser tabs — on every Cloudflare PoP (330+ cities). A request hits the nearest PoP, the isolate spins up in under 5 ms (no cold-start tax), and serves the response from that edge location.

For this server, the first request after the 5-minute cache expires takes one extra hop to GitHub Pages (~50 ms). Everything inside that window is served from edge memory — typically sub-20 ms end-to-end.