Defluff

AI wrote this email. Let's see what they actually wanted.

Reverse the AI behind corporate email: infer the prompt the sender gave their LLM, classify the message, surface the real ask. Zero servers, your keys, your models.

_{A real sales-outreach email, defluffed — the reversed prompt gives it away in one line.}

---

What is this?

Corporate email has become unreadable. Half of every message is AI-generated filler, restated context, and "I hope this finds you well."

Defluff is a browser extension and Outlook add-in that replaces the body of an open email with:

1. Authorship — AI-written, AI-polished, or plainly human. No more always-on "they probably asked an AI" on messages that were clearly written by hand.
2. The reversed prompt — one line naming what the sender probably asked an AI to write ("Pitch our DevOps services and offer a free infrastructure review"). Only shown when AI was actually involved. Often this is all you need.
3. A verdict — Actionable / Response needed / FYI / Noise, with scam-pattern naming for the threats you actually see in 2026: invoice fraud / BEC, phishing, fake recruiter, conference scam, crypto pitch.
4. 3–5 bullets — the actual specifics, preserved verbatim. Returned in the email's own language — French email gets French bullets, German gets German.

"Show original" is one click away if you want the full email back.

The interesting part isn't the summarization — that's just an LLM call with a restrictive prompt. The interesting part is the trust model: there is no Defluff backend. Your email body and your API key never leave your browser for any infrastructure we operate. Each De-Fluff click goes straight from your browser to the LLM provider you picked. Zero retention isn't a policy, it's the architecture.

---

How it works

1. Install Chrome/Edge/Firefox extension, or Outlook Office.js add-in.

2. Configure Click the icon, pick a provider, paste your key. That's it.

3. Defluff Click the button on any email. Prose collapses into bullets.

---

Install

Available now

Chrome Web Store — browser extension for Gmail / Outlook Web / LinkedIn Messaging	Add to Chrome →
Claude Code plugin — Anthropic agent surface	claude plugin marketplace add FinAegis/defluff claude plugin install defluff@finaegis
ClawHub — Openclaw skill	openclaw skills install defluff

Coming soon (pending live-testing — track in #7)

• Edge Add-ons — same build
• Firefox AMO — same build
• Microsoft AppSource — Outlook add-in (desktop + web)

For developers right now

git clone https://github.com/FinAegis/defluff.git
cd defluff
pnpm install
pnpm --filter @defluff/extension build

Then load apps/extension/dist as an unpacked extension (chrome://extensions → Developer mode → Load unpacked). Click the Defluff icon in the toolbar to open settings.

For the Outlook add-in:

pnpm --filter @defluff/outlook-addin dev

Then sideload apps/outlook-addin/manifest.xml from Outlook Web → Get Add-ins → My add-ins → Custom Addins.

---

Providers

Provider	Default model	How it connects
Anthropic Claude	claude-opus-4-7	Messages API with the browser-direct header
OpenAI	gpt-5.4	Chat Completions with max_completion_tokens — works on the reasoning models (o-series / GPT-5) and older chat models alike
Google Gemini	gemini-2.5-pro	Key as query param, native CORS. (Preview tier gemini-3.1-pro-preview available via model override.)
OpenAI-compatible	llama3	Ollama, LM Studio, LiteLLM, or any OpenAI-compatible endpoint. Setup ↓

All four share one extraction prompt. Swap providers anytime from the options page.

A note on the defaults. Defluff defaults to each provider's current frontier model rather than a mini/haiku/flash tier. Scam detection and AI-authorship classification are judgment tasks — smaller models regress meaningfully (they mark obvious AI-recruiter pitches as "legit" because they lean literal). A typical email is ~500-1500 input tokens + ~200 output, so per-click cost lands in the ~$0.01-0.05 range. If that's too much for your usage pattern, override the model to a cheaper tier (claude-haiku-4-5, gpt-4o-mini, gemini-2.5-flash) from the options page — the prompt still handles them, just with less aggression on edge cases.

Running against a local model

Point Defluff at an Ollama or LM Studio instance and get zero per-call cost, full privacy, and no rate limits:

1. Install and run Ollama. brew install ollama (or equivalent) then ollama pull llama3 and ollama serve.

2. Allow the extension origin to call Ollama. Ollama's default CORS config blocks browser requests. Start Ollama with:

   OLLAMA_ORIGINS='chrome-extension://*,moz-extension://*' ollama serve

   Or the more permissive OLLAMA_ORIGINS='*' if you're not fussy. On macOS with the Ollama app, launchctl setenv OLLAMA_ORIGINS '*' then restart the app.

3. Configure Defluff. Options page → Provider: OpenAI-compatible → Base URL: http://localhost:11434/v1 → Model: llama3 (or whatever you pulled) → leave API key blank → Save.

LM Studio is similar: start its local server, point the base URL at http://localhost:1234/v1. LiteLLM and other OpenAI-compatible gateways work the same way.

---

How it compares

	Defluff	Shortwave	Superhuman AI	Grammarly
Your email reaches their servers	Never	Yes	Yes	Yes
Stores your content	Never	Yes	Yes	Yes
Needs an account	No	Yes	Yes	Yes
Monthly cost	$0*	$12	$40	$12
Open source	Yes	No	No	No
Works with local LLMs	Yes (Ollama)	No	No	No

_{* Plus whatever you pay your chosen LLM provider. Typical heavy usage: under $1/month on Claude Haiku or GPT-4o mini. $0 on Ollama.}

---

Privacy

• Your email never touches FinAegis infrastructure. There is no FinAegis infrastructure.
• API keys are stored in chrome.storage.sync (browser) or Office.context.roamingSettings (Outlook) — encrypted at rest, per-user, synced by the platform.
• Extension permissions are host-scoped: mail.google.com, outlook.office.com, and the exact LLM endpoints you'll use. Never <all_urls>.
• Zero analytics, zero telemetry, zero accounts.

Full threat model: SECURITY.md. Reporting: private advisories via GitHub Security.

---

Architecture

  your browser / Outlook webview
  ─────────────────────────────────────────────────
  [De-Fluff button] ──► [email text] ──► [LLM provider]
                                              ↑
                                 api.anthropic.com /
                                 api.openai.com /
                                 generativelanguage.googleapis.com /
                                 your-ollama-endpoint

A pnpm monorepo. TypeScript end to end. Vite for bundling.

Package	What it is
packages/core	Extraction prompt, 4 provider adapters, bullet parser, typed errors. Pure TS, no platform APIs.
apps/extension	MV3 extension for Chrome/Edge/Firefox. React options page, Shadow-DOM-hosted content UI so host-page CSS can't bleed in.
apps/outlook-addin	Office.js add-in. Plain-TS task pane, zero runtime deps.
apps/claude-skill	Claude Code plugin (SKILL.md + manifest). Distributed via the FinAegis marketplace at the repo root — /plugin install defluff@finaegis.
apps/openclaw-skill	A single SKILL.md (markdown + YAML) for the Openclaw agent platform — published to ClawHub.

Provider adapters use plain fetch() — no SDKs in the bundle — to keep client bundles small and the audit surface minimal.

---

Contributing

See CONTRIBUTING.md for setup, the architectural constraints that are non-negotiable (no backend, host-scoped permissions, verbatim extraction prompt), and commit conventions.

The most useful contributions right now are the launch-blocker issues — real-world Gmail/Outlook DOM validation and a CORS smoke test against each provider from a live Outlook task pane.

---

License & support

MIT. Free, forever. Fork it, audit it, ship your own.

If Defluff saves you an hour a week and you'd like to say thanks: ☕ buymeacoffee.com/finaegis. Entirely optional — the product is and always will be free.

_{Built with skepticism by FinAegis.}