GitHub - omerakben/ScribeX

AI-powered academic writing workspace built on Mercury diffusion models

Live Demo · Quick Start · Features · Architecture

ScribeX combines a manuscript editor, writing-mode automation, citation search, multi-format export, humanizer pipeline, AI detection, and structured review workflows into one focused tool for researchers who need to move from rough draft to submission-ready writing with less friction.

Why This Tool Exists

Academic writing is usually split across multiple tools — one for drafting, another for references, another for revision, and yet another for AI prompting. ScribeX pulls those loops into a single interface:

Draft and revise inside a TipTap editor with slash-command AI actions
15 writing commands — generate, expand, simplify, academic tone, outline, counter-argument, evidence, transitions, abstract, rewrite, diffusion draft, mermaid diagrams, summarize, continue, and more
Floating AI actions on text selection — 10 quick actions + 7-mode ribbon panel for rewrite, stylize, fix, custom, tone, humanize, and detect
Writing tools suite — synonyms (5 variants), stylize (8 styles), fix (copy editor), custom instructions, tone analysis
Humanize AI text with few-shot learning pipeline (456-entry dataset, 3-tier UX)
AI detection with sentence-level scoring and color-coded breakdown
Citation search via Semantic Scholar with 6 citation style formats (APA-7, MLA-9, Chicago-17, IEEE, Harvard, Vancouver)
Structured manuscript review with scored category feedback via JSON schema output
Export to 6 formats — PDF, Word (DOCX), Markdown, HTML, BibTeX, and LaTeX
Dark mode with full oklch color scale inversion across all 3 design scales
Local-first persistence with content-hash autosave, per-paper chat, auto-naming, and localStorage hydration
Diffusion drafting — visualize Mercury's parallel denoising process as your draft takes shape

Screenshots

Hero	Editor Commands	Capabilities

Quick Start

Prerequisites

Node.js 20+
pnpm 9+
An Inception Labs API key for Mercury

Setup

git clone https://github.com/omerakben/ScribeX.git
cd ScribeX
pnpm install
cp .env.example .env.local

Edit .env.local with your API key:

INCEPTION_API_KEY=your_inception_api_key_here   # Required
SEMANTIC_SCHOLAR_API_KEY=                        # Optional — increases rate limits
NEXT_PUBLIC_APP_URL=http://localhost:3000        # Use your Vercel URL in production
NEXT_PUBLIC_JOIN_CODE=                           # Leave empty to bypass access gate

pnpm dev

Open http://localhost:3000.

Tech Stack

Layer	Technology
Framework	Next.js 16 (App Router)
Editor	TipTap v3 + ProseMirror
AI Models	Mercury 2 (reasoning) + Mercury Edit (FIM/apply)
State	Zustand 5 with localStorage persistence
Styling	Tailwind CSS v4, shadcn/ui, Radix primitives
Citations	Semantic Scholar API
Export	html2pdf.js, docx, turndown, marked
Animations	Framer Motion + CSS keyframes
E2E Tests	Playwright
Fonts	Manrope (UI), Newsreader (editor), IBM Plex Mono (code)

Architecture

User ─→ EditorCanvas / AIPanel / FloatingMenu
              │
              ├─→ Mercury Client (src/lib/mercury/client.ts)
              │         │
              │         └─→ POST /api/mercury ──→ Inception Labs API
              │              (chat / apply / fim / edit)
              │
              ├─→ POST /api/humanize ──→ Mercury-2 (few-shot humanization)
              │
              ├─→ POST /api/detect ──→ Heuristic AI detection
              │
              ├─→ GET /api/citations ──→ Semantic Scholar API
              │
              └─→ Zustand Stores ←──→ localStorage
                   (scribex:editor / scribex:dashboard)

Browser never calls Inception directly — all AI requests proxy through /api/mercury which attaches the server-side INCEPTION_API_KEY
Middleware (src/proxy.ts) applies CSRF protection and rate limiting (60 req/min per IP) to all /api/* routes
Mercury Client routes writing modes to the correct model: mercury-2 for generative work, mercury-edit for autocomplete/FIM and short edits
Temperature engineering — 22-action temperature map across 5 tiers (0.0 deterministic to 0.9 high creativity) with dynamic token caps based on input length

Feature Status

Feature	Status
Mercury streaming + reasoning controls	Done
Diffusion drafting with denoising overlay	Done
Quick edit / deep rewrite modes	Done
FIM ghost text with word-by-word acceptance	Done
Multi-alternative ghost text (5 cached, Arrow cycle)	Done
Structured manuscript review (JSON schema)	Done
Citation search + style-aware insertion	Done
15 slash commands with model routing	Done
7 paper templates (IMRAD, lit review, thesis, etc.)	Done
Floating menu (10 AI actions) + ribbon panel	Done
Writing tools: synonyms, stylize, fix, custom, tone	Done
Humanizer pipeline (few-shot, 456-entry dataset)	Done
AI detection with sentence-level breakdown	Done
Change block parser + diff cards in AI chat	Done
Document stats + readability badge (Flesch score)	Done
Content-hash autosave + auto-naming	Done
Per-paper chat histories + prompt history (50)	Done
Dark mode (full oklch color scale inversion)	Done
Keyboard shortcuts (5 AI bindings)	Done
Temperature engineering (22-action map, 5 tiers)	Done
HTML sanitization pipeline (5-stage)	Done
Local persistence + autosave	Done
Export: PDF, DOCX, Markdown, HTML, BibTeX, LaTeX	Done
Math (KaTeX) + Mermaid diagram support	Done
Superscript / Subscript	Done
CSRF protection + rate limiting middleware	Done
E2E test suite (10 Playwright specs)	Done
Join-code access gate	Done

Known Constraints

Persistence is browser localStorage only — no server-backed database
Access gating uses a client-side join code, not full authentication
AI detection uses heuristic analysis — production swap to Pangram/GPTZero planned
Plan/usage tier constants exist in code but are not enforced server-side
Rate limiting is in-memory (not suitable for multi-region deployment without Redis)

Usage Examples

Stream a Mercury chat completion

import { streamChatCompletion } from "@/lib/mercury/client";

let output = "";

await streamChatCompletion(
  [{ role: "user", content: "Draft a concise IMRAD introduction on retrieval-augmented generation." }],
  {
    reasoningEffort: "high",
    diffusing: false,
    onChunk: (chunk) => { output += chunk; },
    onDone: () => console.log("done", output),
    onError: (error) => console.error(error),
  }
);

Request structured review output

import { structuredChatCompletion } from "@/lib/mercury/client";
import { REVIEW_JSON_SCHEMA } from "@/lib/constants";

const review = await structuredChatCompletion<{
  categories: { label: string; score: number; feedback: string }[];
}>(
  [{ role: "user", content: "Review this manuscript for argument quality and methodological clarity." }],
  REVIEW_JSON_SCHEMA,
  { reasoningEffort: "medium" }
);

Apply targeted edit

import { applyEdit } from "@/lib/mercury/client";

const revised = await applyEdit(
  "This paragraph has dense and repetitive phrasing.",
  "Simplify while preserving academic tone and factual meaning."
);

FIM autocomplete

import { fimCompletion } from "@/lib/mercury/client";

const suggestion = await fimCompletion(
  "The results demonstrate that",
  "under strict latency budgets."
);

Humanize AI-generated text

import { humanizeText } from "@/lib/mercury/client";

const { alternatives } = await humanizeText(
  "AI-generated text here",
  { count: 4 }
);

Detect AI-written content

import { detectAI } from "@/lib/detection/client";

const { score, label, sentences } = await detectAI(
  "Text to analyze for AI patterns."
);
// label: "human" | "mixed" | "ai"
// sentences: per-sentence scores with color thresholds

Keyboard Shortcuts

Shortcut	Action
`Cmd/Ctrl+Shift+R`	Rewrite selection
`Cmd/Ctrl+Shift+H`	Humanize selection
`Cmd/Ctrl+Shift+F`	Fix grammar
`Cmd/Ctrl+Shift+Y`	Stylize selection
`Cmd/Ctrl+Shift+D`	Detect AI
`Tab`	Accept next ghost text word
`Cmd/Ctrl+Enter`	Accept all ghost text
`Arrow Up/Down`	Cycle ghost text alternatives

API Routes

`POST /api/mercury`

Server-side proxy to Inception API (https://api.inceptionlabs.ai/v1). API keys never reach the browser.

{
  "endpoint": "chat | apply | fim | edit",
  "model": "mercury-2 | mercury-edit",
  "...payload": "endpoint-specific fields"
}

Endpoint mapping: chat → /chat/completions, apply → /apply/completions, fim → /fim/completions, edit → /edit/completions

`POST /api/humanize`

Few-shot humanization pipeline. Assembles examples from a 456-entry dataset server-side, calls Mercury-2.

{
  "text": "AI-generated text to humanize",
  "context": "optional surrounding context",
  "count": 4,
  "action": "generate | generate_one",
  "existing": ["previously generated alternatives for dedup"],
  "temperature": 0.9
}

Returns { alternatives: string[] } for generate or { alternative: string } for generate_one.

`POST /api/detect`

Heuristic AI detection with per-sentence scoring.

{ "text": "Text to analyze (10 chars min, 20K max)" }

Returns { score: number, label: "human" | "mixed" | "ai", sentences: [{ text, score }] }.

`GET /api/citations`

Queries Semantic Scholar and normalizes results into ScribeX citation format.

Param	Required	Default
`q`	Yes	—
`limit`	No	`10`
`offset`	No	`0`

Export Formats

Format	Engine	Notes
PDF	html2pdf.js	Inline CSS with hex colors (oklch incompatible with html2canvas); renders in hidden iframe
DOCX	docx	Full DOM-to-docx with math (Cambria Math), tables, lists, TOC, header/footer
Markdown	turndown	Custom rules for math, mermaid, super/subscript, GFM tables
HTML	Custom	Standalone document with embedded CSS, Google Fonts, KaTeX CDN, print media queries
BibTeX	Custom	Entry type detection, author formatting, double-braced titles
LaTeX	Custom	HTML-to-LaTeX via DOMParser; 12-package preamble; math passthrough

Development

pnpm dev          # Start dev server
pnpm build        # Production build
pnpm lint         # ESLint
npx tsc --noEmit  # Type-check

E2E Tests

# Start dev server first, then in another terminal:
npx playwright test                           # All specs
npx playwright test e2e/04-editor-core.spec.ts  # Single spec
npx playwright test --headed                  # With visible browser
npx playwright show-report e2e/report         # View report

10 spec files covering: landing page, auth gate, dashboard, editor core, slash commands, AI features, citations, autocomplete, persistence, and editor extensions.

Project Structure

src/
├── app/                    # Next.js App Router
│   ├── api/mercury/        # Mercury API proxy route
│   ├── api/citations/      # Semantic Scholar proxy route
│   ├── api/humanize/       # Humanizer pipeline route
│   ├── api/detect/         # AI detection route
│   ├── dashboard/          # Paper management pages
│   └── editor/[id]/        # Editor (dynamic route)
├── components/
│   ├── editor/             # TipTap editor, AI panel, toolbar, slash menu,
│   │                       #   floating-menu, floating-ribbon, change-diff-card,
│   │                       #   dark-mode-toggle, document-stats, readability-badge,
│   │                       #   tone-analysis-card, humanizer-panel, ai-detection-badge
│   ├── dashboard/          # Sidebar, paper cards, new-paper dialog
│   ├── export/             # Export format dialog
│   ├── landing/            # Marketing page sections
│   ├── shared/             # JoinGate, ToasterProvider, DarkModeProvider
│   └── ui/                 # shadcn/ui primitives (direct imports)
├── data/                   # humanizer-dataset.json (456 entries, server-only)
├── hooks/                  # useHydration
├── lib/
│   ├── constants/          # System prompts, commands, templates, temperatures (22-action map)
│   ├── detection/          # AI detection: heuristics, client
│   ├── export/             # 6 format handlers + utils
│   ├── extensions/         # ghost-text (multi-alt), mermaid-block,
│   │                       #   keyboard-shortcuts
│   ├── humanizer/          # Dataset loader, pipeline, types
│   ├── mercury/            # Mercury API client wrapper
│   ├── prompts/            # 37 prompt files: commands/, synonyms/, stylize/,
│   │                       #   fix/, custom/, humanize/, assistant/, document/, system/
│   ├── storage/            # localStorage helpers (SSR-safe)
│   ├── store/              # Zustand stores (editor + dashboard)
│   ├── types/              # TypeScript types
│   └── utils/              # cn, markdown-to-html, change-block-parser,
│                           #   readability, content-hash, sanitize-html,
│                           #   selection-markers, context-window
├── proxy.ts                # CSRF + rate limiting for /api/*
└── types/                  # Module declarations (html2pdf.js)

Why Mercury and Diffusion Matter

ScribeX wraps Inception Labs' Mercury model family using their OpenAI-compatible API surface.

Based on Inception's official materials:

Mercury 2 is a diffusion-based reasoning model focused on high-speed production inference
The Mercury 2 launch post describes parallel refinement (instead of sequential token decoding), highlighting latency-oriented usage for coding, agentic pipelines, and real-time interaction
Inception's platform docs describe OpenAI-compatible endpoints, 128K chat context, reasoning effort controls, structured outputs, and editing endpoints (fim, apply, edit)

ScribeX's value is the workflow layer on top: editor-native controls, writing-mode routing, diffusion UI, review schema integration, and citation-aware UX.

Acknowledgements

Inception Labs — Mercury and Mercury 2 models, API platform
TUEL AI — Project context and publication track
Built by Omer Akben — GitHub · Live Demo

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 16 Commits
docs		docs
e2e		e2e
public		public
src		src
.env.example		.env.example
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
ScribeX_PRD.md		ScribeX_PRD.md
TODO.md		TODO.md
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package.json		package.json
playwright.config.ts		playwright.config.ts
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json

Folders and files

Latest commit

History

Repository files navigation

Why This Tool Exists

Screenshots

Quick Start

Prerequisites

Setup

Tech Stack

Architecture

Feature Status

Known Constraints

Usage Examples

Stream a Mercury chat completion

Request structured review output

Apply targeted edit

FIM autocomplete

Humanize AI-generated text

Detect AI-written content

Keyboard Shortcuts

API Routes

POST /api/mercury

POST /api/humanize

POST /api/detect

GET /api/citations

Export Formats

Development

E2E Tests

Project Structure

Why Mercury and Diffusion Matter

Acknowledgements

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

`POST /api/mercury`

`POST /api/humanize`

`POST /api/detect`

`GET /api/citations`

Packages