MCP Toolkit 🧰

A tiny, free, open-source plugin for both Claude and ChatGPT.

It's a single MCP (Model Context Protocol) server. Because Claude and ChatGPT both now speak MCP, the same server plugs into both — you build it once and one public URL works everywhere.

Tools it adds to the AI

The theme: things AI is confidently bad at. Every tool here computes rather than guesses.

Tool	What it does	Why it helps
get_current_time	Real current time in any timezone	LLMs don't actually know "now"
calculate	Precise arithmetic (+ - * / % ^, parentheses)	LLMs make math mistakes
word_count	Exact word / character / sentence stats	LLMs can't reliably count
test_regex	Actually runs your regex against test strings — real matches, positions, capture groups	LLMs guess regex behaviour and are often wrong
diff_text	Exact line-by-line diff with true added/removed counts	LLMs miscount and invent changes when comparing by eye
estimate_tokens	Exact char/word counts + an approximate token range, and cost if you supply your rate	Sizing prompts and budgeting spend
validate_data	Validates JSON/YAML with a real parser, pinpointing the error line and column	LLMs guess at syntax errors
decode_jwt	Decodes header + payload, renders exp/iat as real dates, flags expiry	Inspecting tokens without a website

All are read-only — they can't change or delete anything.

Honest limits (deliberately stated)

• estimate_tokens is an estimate, not a BPE tokenizer (~4 chars/token heuristic, typically ±10–20% on English prose; code and non-English text differ). The character and word counts are exact. It never hardcodes model prices — you pass your own rate, because prices change.
• decode_jwt decodes; it does not verify. Verifying a signature needs the key. Anyone can read a JWT's contents — only the key holder can prove it's authentic. Don't paste live production tokens into any third-party tool, this one included.
• diff_text caps at 300 lines per side — the diff is O(n×m) and the free tier allows 10ms CPU per request.

---

The one thing to understand first

GitHub stores your code for free, but it can't run a server. (GitHub Pages only serves static files.) To be added to live Claude/ChatGPT, your server needs a public https:// URL that's actually running. So we host the code on GitHub and run it on Cloudflare Workers' free tier.

   Your code on GitHub  ──deploy──▶  Cloudflare Workers (free)  ──▶  https://mcp-toolkit.<you>.workers.dev/mcp
        (free)                            (runs 24/7)                          │
                                                                               ├──▶ paste into Claude
                                                                               └──▶ paste into ChatGPT

Cloudflare Workers' free tier is 100,000 requests/day — plenty for personal use, and it never sleeps.

---

What you need (all free)

• Node.js 18+ installed
• A free Cloudflare account
• A free GitHub account (to publish the code)

---

Step 1 — Get the code & run it locally

# from the mcp-toolkit folder
npm install
npm run dev

You'll see it running at http://localhost:8787. Open http://localhost:8787/ in a browser — you should see a "✔ live" page. (Your local MCP endpoint is http://localhost:8787/mcp.)

Tip: to sanity-check the code before deploying, run npm run type-check.

---

Step 2 — Deploy it (get your public URL)

npx wrangler deploy

The first time, Wrangler opens a browser to log into Cloudflare. When it finishes it prints your live URL:

https://mcp-toolkit.<your-account>.workers.dev

Your connector URL is that address with /mcp on the end:

https://mcp-toolkit.<your-account>.workers.dev/mcp

Open the base URL in a browser to confirm you see the "✔ live" page. Keep the /mcp URL handy — you'll paste it into Claude and ChatGPT next.

---

Step 3 — Put the code on GitHub (make it public & free)

git init
git add .
git commit -m "MCP Toolkit: free MCP plugin for Claude & ChatGPT"
git branch -M main
git remote add origin https://github.com/<your-username>/mcp-toolkit.git
git push -u origin main

That's it — the code is now public and anyone can use it for free (MIT license).

Optional — auto-deploy on every push: this repo includes .github/workflows/deploy.yml. Add a CLOUDFLARE_API_TOKEN secret in your repo (Settings → Secrets and variables → Actions), and every git push will redeploy automatically. (Details are in the workflow file's comments.)

---

Step 4 — Add it to Claude

Works on Free, Pro, Max, Team, and Enterprise (Free is limited to one custom connector).

1. Open claude.ai → your profile/Settings → Connectors (it may be under Customize → Connectors).
2. Click Add → Add custom connector.
3. Paste your URL: https://mcp-toolkit.<your-account>.workers.dev/mcp
4. Click Add. (Leave the OAuth fields blank — this server needs no login.)
5. In any chat, click the + menu → Connectors, enable MCP Toolkit, then ask: "What time is it right now in Asia/Kolkata?" or "Use calculate: (349 * 12.5) / 100".

Works the same in Claude Desktop (Settings → Connectors → Add custom connector). On Team/Enterprise, only an organization Owner can add it.

---

Step 5 — Add it to ChatGPT

ChatGPT reaches custom MCP tools through Developer Mode (a beta). Requirements: Plus, Pro, Business, Enterprise, or Edu, on ChatGPT in a desktop web browser (not the mobile app).

1. ChatGPT (desktop web) → Settings → Connectors → Advanced → turn on Developer mode. (On Business/Enterprise, an admin may need to enable this first.)
2. Back in Connectors, click Create / Add. Fill in:
   • Name: MCP Toolkit
   • Description: Current time, precise math, and text stats
   • URL: https://mcp-toolkit.<your-account>.workers.dev/mcp
   • Authentication: None
3. Save. In a chat, open the + / tools menu, enable MCP Toolkit, and ask the same kind of question.

Why Developer Mode? ChatGPT's older "Deep Research connectors" only accept servers that expose exactly two tools (search + fetch). Developer Mode lifts that restriction so any MCP tools work in normal chat. If your plan can't use Developer Mode, this general-purpose toolkit won't fit the Deep-Research path — but it works fully in Claude.

---

Add your own tool (extend it)

Open src/index.ts and copy the pattern inside init():

this.server.registerTool(
  "reverse_text",
  {
    title: "Reverse text",
    description: "Reverse a string of text.",
    inputSchema: { text: z.string() },          // a plain object of zod fields
    annotations: { readOnlyHint: true },          // read-only = no confirmation prompt
  },
  async ({ text }) => ({
    content: [{ type: "text", text: [...text].reverse().join("") }],
  }),
);

Then npx wrangler deploy again (or just git push if you set up auto-deploy). Reconnect in Claude/ChatGPT to pick up the new tool.

---

Troubleshooting

Problem	Fix
Connector won't connect	Make sure you used the /mcp URL and that opening the base URL in a browser shows the "✔ live" page. The server must be publicly reachable (Cloudflare's is).
npx wrangler deploy fails to log in	Run npx wrangler login once, then retry.
ChatGPT has no "Developer mode"	It needs a Plus+ plan on desktop web; on Business/Enterprise an admin must enable it.
Tools don't appear in a chat	Open the + / connectors menu in that chat and toggle the connector on.
Don't put secrets in the URL	Never add ?token=... to the URL — it's blocked by the MCP spec. Use proper auth if you need it.

---

How it's built (for the curious)

• Runtime: Cloudflare Workers (serverless, free tier, never sleeps)
• Library: agents' McpAgent + the official @modelcontextprotocol/sdk
• Transports served: /mcp (Streamable HTTP, recommended) and /sse (legacy, for older clients)
• Auth: none (authless) — anyone with the URL can use the read-only tools

License

MIT — free for anyone to use, modify, and share. Replace <Your Name> in LICENSE.