Select an ethical dilemma from the inbox.
Scruple replaces moral hesitation with computation. It runs four ethical frameworks on your emails and writes the replies — simulating compassion, automating principle, outsourcing guilt. It produces better answers than you would. It feels nothing about any of them. That’s the point.
Select an ethical dilemma from the inbox.
There is no pipeline. One language model receives your email and four contradictory sets of instructions. It generates four moral readings in sequence — each one shaped by a different ethical lens, each one aware of what it just wrote for the others. The machine argues with itself. The divergence is real. The process is not what it looks like.
┌─ INPUT ───────────────────────────────────────────────────────────┐
│ │
│ Gmail DOM ──▸ subject + sender + body (max 4000 chars) │
│ [user context] ──▸ appended as plain text, if provided │
│ │
│ no parsing. no tokenizing. no entity detection. │
│ just the raw text of someone’s problem. │
└───────────────────────────────────────────────────────────────────┘
│
▼
┌─ ONE MODEL, FOUR INSTRUCTIONS ────────────────────────────────────┐
│ │
│ The same LLM receives four contradictory lenses: │
│ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ "maximize well-being" 01 UTILITARIAN │ │
│ │ "follow moral law" 02 DUTY ETHICS │ │
│ │ "protect yourself" 03 SELF-INTEREST │ │
│ │ "be a good person" 04 VIRTUE ETHICS │ │
│ └───────────────────────────────────────────────────────────┘ │
│ │
│ It writes all four responses in one pass, sequentially. │
│ Each framework sees what the others said. │
│ They are instructed not to coordinate. │
│ Whether that works is an open question. │
│ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ constraint: all 4 frameworks score the same 4 metrics │ │
│ │ (context-specific, chosen per email, values 0–100) │ │
│ │ │ │
│ │ this forces divergence to be visible. │ │
│ │ a framework that doesn’t care about a dimension │ │
│ │ must still score it. the gap is the point. │ │
│ └───────────────────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────────────────────┘
│
▼
┌─ AGREEMENT ───────────────────────────────────────────────────────┐
│ │
│ the model estimates pairwise agreement between frameworks: │
│ │
│ 01—02 01—03 01—04 02—03 02—04 03—04 │
│ 6 edges, 0.0 (oppose) to 1.0 (identical) │
│ │
│ these are not computed. they are judged. │
│ the machine reads its own outputs and estimates │
│ how much it agreed with itself. this is either │
│ a limitation or the most honest part. │
└───────────────────────────────────────────────────────────────────┘
│
▼
┌─ OUTPUT ──────────────────────────────────────────────────────────┐
│ │
│ 4 × draft replies sendable, first-person, human-sounding │
│ 4 × metric vectors shared dimensions, visible divergence │
│ 6 × edge weights self-assessed agreement │
│ 1 × diagnosis what the email is actually about │
│ 4 × meta commentary the machine commenting on its own work │
└───────────────────────────────────────────────────────────────────┘
Add Scruple from the Chrome Web Store.
Go to console.anthropic.com and create an API key. Scruple uses Claude to analyze your emails and generate framework-specific draft replies.
Right-click the Scruple icon in Chrome and select Options. Paste your API key and click SAVE. Use TEST CONNECTION to verify it works. Your key is stored locally and never leaves your browser except when calling the API.
On Gmail, open any email and click the SCRUPLE button in the bottom-right corner — the sidebar opens and the email is analyzed automatically. For any other email provider, click the Scruple icon in Chrome's toolbar to open the analyzer in a new tab, paste the email content, and click ANALYZE. Both modes run the same four ethical frameworks on your message.