Irrational MCP Server

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.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:

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

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

{
  "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

npm install --ignore-scripts   # we don't use image optimization, so skip sharp's native build
npm run dev                    # http://localhost:4321
npm 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.