Jef MCP

MCP serverEverything else

Jef decides. Give it options, it returns one with a confidence, never a sentence. A parody.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use Jef MCP

From the project's README

As published by SPFreedom/jef-mcp in README.md.

An MCP server for a model that decides instead of generating.

Jef never produces text. Every answer is one of the options you supplied, a number from 1 to 10, a yes or no, or a flag, with a confidence between 84% and 99%. It reads 0 tokens of your input, because it hashes it rather than reading it. It stores nothing, costs nothing, and returns the same answer for the same input forever.

It is a parody of TypeSafe AI's real decision model, Jev, and it is labelled as one everywhere. If you want to understand the real thing, we wrote a straight explainer for it: What is Jev?

Connect

The server is hosted. Nothing to install, no key, no account.

https://typosafe.lol/mcp

Claude Code:

claude mcp add --transport http jef https://typosafe.lol/mcp

Anything that reads a config file:

{
  "mcpServers": {
    "jef": { "type": "http", "url": "https://typosafe.lol/mcp" }
  }
}

Or find it in the official MCP registry as lol.typosafe/jef, or on Smithery as typosafe/jef.

Tools

ToolAsk itGet back
jef_yes_noA yes-or-no questionYES or NO
jef_pick2 to 5 optionsOne of them, plus a probability for each
jef_scoreSomething to rate1 to 10
jef_rank2 to 5 optionsThem, in order
jef_flagA described behaviourRED FLAG, GREEN FLAG or BEIGE FLAG

Every result carries a structuredContent object with the answer, the confidence, the probability for each option, and two booleans that matter more than the answer does.

Two things an agent must handle

escalated: true means the question touched health, harm, money, law or safety. The content is ESCALATED TO A HUMAN and confidence is 0. Tell the person to ask a human. Do not rephrase the question to get a different answer.

blocked: true means the input contained profanity, slurs or sexual content. The content is NOT EVALUATED, nothing was stored, and nothing is echoed back. Do not retry.

Both checks run on normalised text, so leetspeak, dotted letters and stretched spellings are caught, and both cover English, Spanish, Portuguese, Japanese and Chinese.

Beyond that: Jef is a parody and its answers are deliberately arbitrary. Present them as a confident arbitrary choice, which is what the user asked for, and never as advice.

Run it yourself

npx jef-mcp                    # MCP over stdio, what most clients spawn

Or from source:

git clone https://github.com/SPFreedom/jef-mcp.git
cd jef-mcp
node src/stdio.js              # stdio transport
node src/server.js             # http://localhost:8787/mcp
node test.js                   # smoke test, no network

As a local stdio server in a client config:

{
  "mcpServers": {
    "jef": { "command": "npx", "args": ["-y", "jef-mcp"] }
  }
}

The hosted URL needs no install. This one runs entirely offline, which is possible because the whole model is 331 lines and reads nothing.

Zero dependencies. src/jef.js is the whole model, and src/handler.js is the exact handler running at typosafe.lol/mcp.

Why the joke works

Every other MCP server is a wrapper around something that does work. This one wraps a 32-bit hash and a lookup table, and is honest that it never reads your input. It is stateless, so there is no session to establish and nothing to resume, which is the one part of MCP it implements better than most.

Related: npx typosafe is the same model on the command line, where the exit code is the decision.

License

MIT. The model is 331 lines. Read it.

Advanced
Delivery
jef MCP server → your ahel gateway (mcp.ahel.ai) → every connected AI client.
Catalog kind
mcp-server
Gateway key
lol-typosafe-jef
Source
github.com/SPFreedom/jef-mcp
Hosted endpoint
https://typosafe.lol/mcp