Irrational

io.github.urbanmorphPublished by GitHub account urbanmorphv0.1.0Updated Oct 25, 2026

Adversarial behavioural-bias engine — audits your decisions for cognitive biases via your own AI.

VerifiedStreamable HTTPWeb executableKnowledge & MemoryAI & ML

Overview

AI-generated overview

A remote MCP server that gives an assistant a 22-bias catalogue and an adversarial prompt for auditing decisions for cognitive biases.

What you can do
Audit a decision for cognitive biases, returning a verdict-first directive the assistant's own model executes to produce the reasoning and a confidence estimate, with each bias tied to the user's words and optionally in another language or as structured JSON (analyze_decision). Browse the 22-bias catalogue, optionally filtered by family (list_biases). Look up the full entry for a single bias by id, including the question that breaks its spell and a citation (get_bias).
When to use it
Useful when you want a second, deliberately contrarian pass over a plan, estimate or draft reasoning, and want the assistant to argue the other side rather than validate it. Also handy for looking up a named bias and its counter-question.
Requirements
A remote streamable HTTP endpoint; no install, API keys, environment variables or headers are needed. The server runs no model itself, so the connected assistant's own model does the reasoning.
Before you install
The server returns a directive that the assistant's model then executes, so the audit's quality depends on that model. The README states no decisions are stored and nothing leaves the client, but the decision text is sent to the remote endpoint to compose the directive.

Installation

In SourceWeft

  1. Open Irrational in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Use with an agent

Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.

Read https://sourceweft.com/mcp/io.github.urbanmorph%2Firrational/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.

Read the install guide the agent follows

Other MCP clients

Add this to your client's mcpServers config.

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

README

Irrational

A field guide to the mistakes your mind makes, and an open, adversarial behavioural-bias engine you can call from your own AI or any MCP agent.

[Irrational: a field guide to the mistakes your mind makes, and an engine that audits your decisions]

[Live] [MCP Registry] [CI] [License: MIT] [PRs welcome]

→ irrational.pages.dev · Field guide · Connect via MCP · Sources

Irrational runs no AI model. It provides the curated adversarial framework, a 22-bias catalogue, and a stateless scaffold, delivered as a public field guide and an open MCP tool, and the reasoning is done by whatever model you or your agent already have. No API keys. No stored decisions. Nothing leaves you.

A frontier model already knows the biases, so the value isn't informational; it's behavioural. A plain LLM is trained to agree, while Irrational has a fixed adversarial contract (assume you're biased, refuse to validate, argue the other side) that an agent can invoke systematically, even on its own draft reasoning.

What an audit looks like

You: "We should rewrite our 6-year-old backend from scratch this year, we've already spent two months on a POC that went well, and I'm confident we can finish in six."

THE CALL: Don't greenlight the big-bang rewrite. Carve off the worst module first, with a hard 3-month checkpoint. Confidence the 6-month plan succeeds: ~25%, mostly sunk cost + the planning fallacy.

THE AUDIT: Those "two months on a POC" are gone either way, but they're doing most of the talking (sunk cost). "About six months" is a best-case with no reference class; rewrites of mature systems routinely run 2–3× (planning fallacy)…

Verdict-first, every bias tied to your own words, grounded in a cited catalogue. See the full worked examples →

Add it to your AI

Listed on the official MCP Registry as io.github.urbanmorph/irrational. Connect any MCP client to the live endpoint, no install, no keys:

bash
# Claude Codeclaude mcp add --transport http irrational https://irrational.pages.dev/mcp/server

…or point any MCP client at https://irrational.pages.dev/mcp/server:

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

Tools: analyze_decision (the adversarial audit) · list_biases · get_bias. Pass language for a reply in another language, or structured: true for parseable JSON.

What's inside

  • Field guide: 22 cognitive biases grouped by Buster Benson's four problems the brain solves, each with the question that breaks its spell and a verified citation.
  • The engine: an MCP server exposing analyze_decision, list_biases, get_bias. analyze_decision returns a verdict-first directive your model executes. The discipline is ours; the reasoning is yours.
  • The web forge: type a decision; it assembles the adversarial prompt in your browser and hands off to your AI. No model runs here.
  • Pen-and-paper audit: the whole method on one printable page.

Stack

Astro (pure static) · Cloudflare Pages · a hand-rolled, dependency-free JSON-RPC MCP endpoint as a Pages Function. No runtime model, no database. Single source of truth: src/data/biases.ts feeds both the web and the MCP.

Develop

bash
npm install --ignore-scripts   # we don't use image optimization, so skip sharp's native buildnpm run dev                    # http://localhost:4321npm test                       # vitest, engine + catalogue (29 tests)npm run build                  # static pages

The MCP endpoint is a Cloudflare Pages Function, run it locally with npx wrangler pages dev dist.

Contributing

Translations, proposed biases, and worked examples are welcome, see CONTRIBUTING.md. This project also publishes a public PDGI scorecard.

Licence

Code is MIT. The bias-catalogue text (definitions, examples, and the curated prose in src/data/) is additionally available under CC-BY-4.0. Every claim is cited, see /sources.

Built by Urban Morph.

Source: README.md at commit bb7f8b4

Tools

3

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.

Parametersmodejudgmentrequiredlanguagereasoningstructured

list_biases

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

Parametersfamily

get_bias

Get the full entry for one bias by id.

Parametersidrequired

Version history

1
  1. v0.1.0LatestSep 16, 2026

Score

100% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 daysRequired25 pts
  • Auth declaredRequired10 pts
  • Tools listedRequired15 pts
  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts