# Thinking Tools 🧠

Five rigorous **reasoning protocols** for Claude & ChatGPT. Free, open-source, and **zero extra credits** —
they run on the AI you're already talking to.

One [MCP](https://modelcontextprotocol.io) server. Works in both Claude and ChatGPT.

| Tool | What it does |
|------|--------------|
| `debate` | Convenes an expert panel **derived from your question**, argues it out (steelman → attack), then delivers a verdict **with the strongest dissent preserved** and what evidence would flip it |
| `red_team` | Steelmans your plan, then hunts hidden assumptions, failure modes, misuse, scaling breakpoints and single points of failure. 12-month pre-mortem. Risks ranked by likelihood × impact, each with the cheapest fix or a falsifying test |
| `audit_argument` | Reconstructs any text's argument, names **quote-anchored** fallacies, biases and rhetoric, shows what evidence is missing — and rates *argument quality* and *conclusion truth* **separately** |
| `threat_model` | Defensive STRIDE review: assets, actors, trust boundaries, data flows → rated threats, each with a **named, located control**. Hardening only, never exploits |
| `check_study` | Extracts design/N/baseline/effect size/CI/p from a study or statistic; flags p-hacking, missing controls, survivorship bias, relative-risk inflation; verdict separates *weak study* from *false claim* |

## How it's free

The server runs **no AI**. Each tool returns an expert-designed **protocol** as plain text; your own
Claude/ChatGPT reads it and does the reasoning in the same turn. No API key, no second bill.

That's also why you **invoke it explicitly** (`"Debate this: ..."`). An MCP server can't intercept your
message before the AI sees it — that's a platform rule, not a missing feature.

## Design principles baked into every protocol

- **Steelman before attacking** — never critique a weaker version than you'd defend
- **Quote before accusing** — no quote, no fallacy claim
- **Never fake consensus** — dissent is preserved, split verdicts are reported as split
- **Invent nothing** — no fabricated stats, sources, or citations; uncertain recall is labelled
- **Separate "bad argument" from "false conclusion"** — a bad argument can have a true conclusion

## Use it

```
Debate this: should we rewrite the backend in Rust?
Red team my plan: [paste plan]
Audit this argument: [paste article]
Threat model: a public API with JWT auth and Postgres
Check this study: "Coffee cuts heart disease 50%, per a study of 40 people"
```

## Deploy your own (free)

```powershell
npm install
npx wrangler deploy      # prints your live URL
```
Your connector URL is that address **+ `/mcp`**.

## Connect it
- **Claude** (any plan): Settings → Connectors → Add custom connector → paste the `/mcp` URL
- **ChatGPT** (Plus+, desktop web): Settings → Connectors → Advanced → Developer mode → add the `/mcp` URL, Auth **None**

> ⚠️ Use the URL ending in **`/mcp`** — the bare address is just a status page.

## License
MIT © 2026 Pranay Mahendrakar — free to use, modify, and share.