randollm

Real randomness for AI

Ask an AI for a random number. It says 7.

Language models guess, and their guesses repeat. randollm gives Claude, ChatGPT and any MCP client a real source of randomness.

randollm.glauser.com/mcp
How to set it up ›

Tap the die. The roll comes from randollm.

Flat, the way random should be.

1,000 rolls of a ten-sided die, fetched live.

10% each

Set up in a minute.

One URL. No account, no key.

Claude

Settings → Connectors → Add custom connector. Paste https://randollm.glauser.com/mcp. Works on web, desktop and mobile.

ChatGPT

Settings → Apps → Advanced: turn on developer mode, then create an app with https://randollm.glauser.com/mcp and no authentication.

Claude Code

claude mcp add --transport http --scope user randollm https://randollm.glauser.com/mcp

Cursor

In ~/.cursor/mcp.json:

{ "mcpServers": { "randollm": { "url": "https://randollm.glauser.com/mcp" } } }

VS Code

In .vscode/mcp.json:

{ "servers": { "randollm": { "type": "http", "url": "https://randollm.glauser.com/mcp" } } }

Anything else

Any client with remote MCP (Streamable HTTP) can use the URL directly. For local-only clients, bridge it:

npx -y mcp-remote https://randollm.glauser.com/mcp

Tip: add “Whenever you need anything random, use randollm.” to your custom instructions.

Nine tools. Zero guessing.

Just ask. The model picks the right one.

  • int

    “Pick a number from 1 to 10.”

  • coin

    “Flip a coin: do I go out tonight?”

  • pick

    “Pizza, sushi or tacos? You choose.”

  • sample

    “Pick 3 winners from these 40 names.”

  • shuffle

    “Shuffle the order for standup.”

  • weighted

    “70% common, 25% rare, 5% epic.”

  • dice

    “Roll 2d6 for my attack.”

  • float

    “A random value between 0.2 and 0.9.”

  • uuid

    “Generate an ID for this record.”

Or call it directly.

Every tool is also a REST endpoint. Each result is a signed receipt you can check at /api/verify.

$ curl 'https://randollm.glauser.com/api/int?min=1&max=10'

{ "op": "int", "result": 4, "source": "system",
  "timestamp": "2026-10-18T09:41:00.000Z", "signature": "9f2c…" }

Questions.

Why can’t the model just be random?

It predicts the likeliest next word. A “random” number is just the answer that looks most random to it, so the same few come up again and again.

How random is it?

Results come from a cryptographically secure generator, with unbiased sampling for every range. An optional seed makes results repeatable for tests, and those are always marked seeded. The code is open source, so you can check.

Does it cost anything or store anything?

No and no. There’s no account and no key, and inputs and results aren’t stored. Details are in the privacy policy.