The Engine · MCP

Wire the bias engine into your agent.

Irrational is an open, stateless MCP server. It runs no model; it returns the adversarial framework, the 22-bias catalogue, and the procedure as a directive your agent's own model executes. No keys, nothing stored.

Listed on the official MCP Registry as io.github.urbanmorph/irrational ↗

Endpoint

https://irrational.pages.dev/mcp/server

Add it to Claude Code

claude mcp add --transport http irrational https://irrational.pages.dev/mcp/server

…or any MCP client

{
  "mcpServers": {
    "irrational": { "type": "http", "url": "https://irrational.pages.dev/mcp/server" }
  }
}

Tools

analyze_decision

Adversarially audit a decision for cognitive biases. Returns a directive YOUR model executes to produce the composed audit (verdict-first). Provide reasoning, not just the conclusion.

list_biases

List the 22-bias catalogue, optionally filtered by family.

get_bias

Get the full entry for one bias by id.

The pattern

An agent calls analyze_decision with a judgment and the reasoning behind it. The tool returns a directive: the adversarial contract, the catalogue, the procedure, and the required verdict-first output shape. The agent's own model produces the audit, as readable prose by default. The discipline is ours; the reasoning is yours.

Need machine-parseable output? Pass structured: true and the audit comes back as a JSON object (bias ids and keys in canonical English) you can store or compare across decisions.

Best used systematically: have an agent audit its own draft recommendation before it finalizes, or let a second agent red-team the first. That external check is where the real debiasing lives.

Any language

Pass an optional language (e.g. "Tamil", "Spanish") and the audit comes back in that language; the bias ids and output keys stay canonical English, so the result is still parseable and trackable. No translation files needed; your model does it in-context.