01 The case for sovereign AI

Some documents shouldn't leave the room.

DREA brings AI to your documents — not your documents to AI. Pick a specialist for legal, clinical, financial, technical, or academic work, and let it think alongside you on the Mac you already own.

Get DREA How it works
For the work that can't be uploaded
Designed in Georgia. Runs entirely on your Mac.
Master Services Agreement p. 14 of 200
8. Confidentiality & Non-Disclosure
The parties agree to maintain the confidentiality of all proprietary information disclosed during the term of this agreement.
Saul-7B reading
Your Journey

You're already using AI.
DREA makes it safe.

You're a professional exploring what AI can do for your work—analyzing contracts, synthesizing research, evaluating strategies. The possibilities are extraordinary. But every step forward comes with a question: where is my data going?

DREA is built for that question.

She's your companion on this journey—a local-first platform that processes documents, runs AI reasoning, and manages your knowledge base right on your machine. And when you want to reach further—to leverage Claude, ChatGPT, or any external AI—DREA acts as a trust layer, sanitizing sensitive information before anything leaves your environment.

You lead the discovery. You direct the research. You make the evaluations. DREA is simply the tool that ensures you can do all of it without compromise.

One platform · every document type · zero uploads
Capabilities

Use any AI.
Compromise nothing.

DREA Desktop combines on-device document processing, local semantic search, and air-gapped AI inference on Apple Silicon—and when you need more, she safely bridges you to external AI without exposing what matters.

01 / 07
The trust proxy

Use Claude or ChatGPT — without surrendering your data.

DREA sanitizes PII and proprietary information before anything leaves your environment. Full audit logging means you always know what was shared.

Trust proxy Live
In Review the contract for client Jane Doe (#84291) signed 03/14/2026…
Out Review the contract for XXXXXXX (#XXXXX) signed XX/XX/XXXX…
02 / 07
Document engine

Multi-format ingestion — everything reads.

PDFs, Word, PowerPoint, Excel, HTML, Markdown, OCR'd images, and video up to 500MB—all processed locally. Drag, drop, done.

Ingest queue 7 formats
PDFmaster_services_agreement.pdf2.4 MB
DOCXq4_strategy_brief.docx812 KB
XLSXvendor_reconciliation.xlsx1.1 MB
MP4deposition_transcript.mp4412 MB
IMGscanned_invoice_0314.jpg3.2 MB
03 / 07
Search

Search that thinks before it answers.

Every query is classified, expanded into semantic variants, and reranked by a dedicated helper LLM. You get grouped evidence cards with relevance notes—not raw excerpts.

Query pipeline Reranked
> "When does the perpetual NDA clause kick in?"
01Sec. 8.2 · perpetual confidentiality term0.94
02Sec. 8.4 · carve-out exclusions0.81
03Exhibit C · confidentiality definitions0.72
04 / 07
Reasoning

Local reasoning, specialist by sector.

Qwen3-4B routes every query; your sector specialist handles deep chain-of-thought reasoning. Both run natively on Apple Silicon—fully offline.

Model handoff On-device
Helper
Qwen3-4B
classify · expand · rerank
↓
Specialist
Saul-7B-Instruct
chain-of-thought · legal-tuned
05 / 07
Orchestration

Agents that run the research.

Specialized agents for documents, queries, system, and workspace operations coordinate autonomously—while staying in your control.

Active agents 4 / 4
Document agent
Query agent
System agent
Workspace agent
06 / 07
Security

Zero-trust, air-gapped by default.

API key auth, path-traversal protection, network scope validation, trust-proxy PII sanitization. Privacy-first security at every layer.

Security posture All checks pass
✓API key authentication
✓Path-traversal protection
✓Network scope validation
✓Trust-proxy PII sanitization
✓Audit log enabled
07 / 07
Operations

Admin dashboard — everything in your sights.

Real-time health monitoring, agent session tracking, log viewer, hook management, and full configuration control—built for operators.

Health Operational
Sessions
12
Latency
214ms
Uptime
99.9%
Experience

See DREA in action.

A short walkthrough of the discovery flow — from blank prompt to cited answer, all on-device.

discovery-walkthrough.mov
Recorded May 2026 2.5×

A short walk through DREA's discovery flow, on a 24 GB MacBook Air.

Under the Hood

Local foundation.
Global reach.

Every core component runs on your machine. When you choose to reach external AI, DREA's trust proxy stands between your data and the outside world.

DREA Desktop
DREA Desktop Tauri + Rust
Search UI Vue 3
Service Manager Rust
DREA API FastAPI :8000
ChromaDB Embedded
llama.cpp LLM :8080
DREA API FastAPI :8000
Admin UI Vue 3 :8001
Qdrant Vector DB :6333
Embeddings BGE GGUF + GME
llama.cpp LLM :8080
Docker / Podman Compose

How Your Data Flows

📄
Documents
🔧
Parse & Chunk
🧮
Embed Locally
🗃
Vector Store
💡
AI-Powered Answers
Choose Your Specialist

One platform.
A specialist for every desk.

Drop in a model trained for your world and let it read, search, and reason across your documents at a depth that turns hours of manual review into minutes. Every specialist runs entirely on your Mac.

What You Need

Runs on the Mac
you already have.

DREA Desktop is built for Apple Silicon and tuned for local, air-gapped AI. Pick the Mac configuration that fits how you work — you can always grow into more capability later.

Start Here
Intelligent Search
vRAM

DREA Lite

Apple Silicon · 16GB Memory
Qwen3-4B + BGE-small Helper LLM routes every query
  • AI-powered query classification, expansion, and reranking
  • Grouped evidence cards with relevance notes
  • Spreadsheet data rendered as formatted tables
  • Safely use Claude, ChatGPT, or Copilot via trust proxy
  • Detect topics, duplicates, and related content
  • Process PDFs, Word, Excel, images, and more
Recommended
Full AI Orchestration
vRAM

DREA

Apple Silicon · 24GB Memory
Qwen3-4B + Your Specialist + External AI Helper + the specialist of your choice + cloud AI
  • Everything in Intelligent Search, plus:
  • Pick a 7–14B specialist tuned for your sector
  • Written answers with step-by-step reasoning
  • Back-and-forth conversations about your docs
  • Multi-LLM orchestration — local and cloud AI working together
  • Add vision model for image and video search
Maximum Power
vRAM

DREA Pro

Apple Silicon · 64GB+ Memory
Full Stack + Larger Specialists Run 32B–70B specialists at full quality
  • Everything in Full AI, plus:
  • Run the largest specialists (32B–70B) at full quality
  • Handle thousands of documents at once
  • Map relationships across your entire library
  • Built for teams and organizations

What you can do at each level

What DREA Can Do 16GBSearch & Discover 24GBFull AI 64GB+Professional
Intelligent Search (Qwen3-4B Helper LLM)
AI query classification, expansion, and reranking ✓ ✓ ✓
Grouped evidence cards with relevance notes ✓ ✓ ✓
Spreadsheet data rendered as formatted tables ✓ ✓ ✓
Safely work with Claude, ChatGPT, or Copilot ✓ ✓ ✓
Automatically summarize, tag, and classify documents ✓ ✓ ✓
Process PDFs, Word, Excel, images, video & more ✓ ✓ ✓
Deep Analysis (Requires a Specialist Model, 24GB+)
Written answers with step-by-step reasoning — ✓ ✓
Back-and-forth conversations about your docs — ✓ ✓
Multi-LLM orchestration (local + cloud AI) — ✓ ✓
Automated multi-step research and analysis — ✓ ✓
Image and video search (with vision model) — ✓ ✓
Map relationships across your entire library — — ✓

With 16GB, DREA is a smart research assistant — Qwen3-4B classifies, expands, and reranks every query, delivers grouped evidence cards, and safely connects to external AI via the trust proxy. At 24GB, DREA becomes a full AI orchestration platform — local reasoning, cloud AI, and multi-step analysis working together through one interface. You start wherever you are and grow when you're ready.

System Requirements

OS macOS 12.0+ (Sonoma or Sequoia recommended)
Chip Apple Silicon
RAM 16GB minimum, 24GB+ recommended
Storage 20GB+ free (models + vector database)
Python 3.10+ (bundled with DREA)
Network Not required (fully offline operation)
Under the Hood

The AI that powers DREA.
All local. All yours.

Every AI model runs entirely on your Mac. No cloud, no accounts, no data leaving your device. Two models handle every search out of the box—two more unlock vision and deep reasoning when you're ready.

Core — Powers Every Search
Core Engine

Qwen3-4B

The Brain Behind Every Search

Qwen3-4B is the intelligence layer that makes DREA's search actually smart. Before your query ever hits the document index, this model classifies your intent, expands your words into semantic variants, and reranks every result by relevance with a reason. It also powers the trust proxy that keeps your data safe when using External AI.

Why this model? At 2.5 GB quantized (Q4_K_M), Qwen3-4B handles six distinct tasks—classify, expand, rerank, summarize, tag, and conversational fallback—with sub-second response times. Built by the Alibaba Qwen Team, it's optimized for structured JSON output, making it ideal for pipeline utility tasks without the overhead of a large reasoning model. It runs on any Apple Silicon Mac, even 16 GB.
What Qwen3-4B does on every search:
Classifies your query as search, conversation, or needs-clarification
Expands your words into 3 semantic variants for better recall
Scores and reranks every result with a relevance reason
Sanitizes queries for External AI with zero data leakage
Always Active Open Source 2.5 GB · Sub-second 6 Tasks

BGE-small GGUF

Document Memory

When you search for something, BGE-small is why DREA finds it—even when you don't use the exact right words. It converts every document into a searchable vector, enabling semantic matching by meaning, not just keywords.

Why this model? At just 35 MB in a single GGUF file, it loads in under 2 seconds and runs on any Apple Silicon Mac with zero slowdown. Built by the Beijing Academy of AI (BAAI), BGE-small is one of the most widely deployed open-source embedding models. DREA ships the Q8-quantized variant—near-lossless quality at a fraction of the original size.
Always Active Open Source 35 MB · 2s Cold Start
Optional — Expand Your Capabilities

GME-Qwen2-VL

Visual Intelligence

Adds image and video understanding. Search across screenshots, charts, slides, and video frames with natural language. Built by Alibaba DAMO Academy. Without it, DREA still searches all text content.

Add Vision 4 GB
Your Specialist

Specialist Model

A reasoner trained for your world

Drop in a model tuned for your sector—legal, clinical, financial, technical, or academic. It writes answers, holds conversations, and reasons over your documents at the depth your work demands. You choose what runs here—or skip it entirely.

Pick by sector
7 specialists, one slot → Saul-7B (Law) · OpenBioLLM (Health) · Qwen2.5-Coder (Engineering) · R1-Distill (Academic) · and more
Also Compatible
Llama Mistral Phi Gemma Any GGUF / MLX model
Why is this optional? DREA already finds and ranks your documents intelligently without a specialist—Qwen3-4B and BGE-small handle that. Adding a specialist unlocks written answers and deep multi-step analysis. The memory it needs depends on which one you pick — most fit a 16 GB Mac, the largest scale to 32 GB+. We give you the choice so DREA works on the hardware you have.
What a specialist unlocks:
Domain-trained reading and reasoning
Back-and-forth conversations about your docs
Step-by-step reasoning you can follow
Automated multi-step research
Optional Swappable Scales 8 GB → 32 GB+ Mac
Where Your Data Lives

The AI models above understand your documents. The databases below remember everything they learn—every meaning, every connection, every visual detail—so search is instant and nothing is ever lost.

Built In

ChromaDB

Your Document Vault

ChromaDB is where DREA stores everything it learns about your documents—meanings, relationships, visual content, and more. When you search and get results in under a second, this is why. It runs inside DREA itself, so there's nothing extra to install, configure, or manage.

Why this database? ChromaDB is embedded directly into DREA—no separate server, no background process, no network connection. It works fully offline and keeps your data entirely on your machine. For personal and single-user workflows, it handles thousands of documents without breaking a sweat. You never have to think about it—it just works.
Always Active Embedded Zero Config
You Choose

Qdrant

Scalable Storage

When you need more—more users, more documents, or a dedicated database server—Qdrant steps in. It's a standalone vector database that separates your storage from the app, giving you the flexibility to scale without limits.

Why is this optional? ChromaDB handles everything for personal use beautifully. Qdrant is there when you outgrow what's built in—team environments, enterprise deployments, or document libraries in the tens of thousands. DREA can switch between them seamlessly, and your data works the same way regardless of which one you choose.
When to consider Qdrant:
Multiple users accessing the same document library
Very large document collections (10,000+)
Separating the database from the application
Server or enterprise deployments
Optional Server-Based Team Ready

Every model and database is open-source, runs offline, and is included in your DREA download. No accounts, no API keys, no subscriptions to third parties. You can inspect, replace, or remove any component at any time.

Get DREA

Your documents are costing you money.
DREA pays for itself.

Stop losing hours digging through files, re-reading reports, and chasing down answers your team already has. DREA Desktop puts everything you know at your fingertips — privately, instantly, on your Apple Silicon Mac. One price. No subscriptions.

One-Time Purchase

DREA Desktop

$ 299

No subscriptions. No cloud fees. No per-seat pricing.

  • ✓ Your sensitive data never touches the cloudClient files, financials, legal documents, medical records — use AI without risking a breach, a leak, or a compliance violation. Everything stays on your Mac.
  • ✓ Get answers from your own documents in secondsStop digging through folders and email chains. Ask a question in plain English and DREA pulls the answer with the exact source — across every file you've ever loaded.
  • ✓ Drop in any file format and start workingPDFs, Word, Excel, PowerPoint, HTML, images, video — 22 formats. No converting, no reformatting, no wasted time getting files ready.
  • ✓ Reclaim hours buried in document reviewDue diligence, contract comparisons, inventory spreadsheets, audit prep — tasks that used to take half a day now take minutes. Ask DREA to find a part number, compare line items, or summarize a 200-page report.
  • ✓ No subscriptions draining your budgetOne purchase, one price. No per-seat fees, no usage caps, no monthly bill that grows with your team. DREA pays for itself the first week you use it.
  • ✓ Use external AI without exposing your dataDREA sanitizes sensitive information before it reaches ChatGPT, Claude, or any outside model — so you get the power of cloud AI without the risk.
  • ✓ Orchestrate local and cloud AI togetherQwen3 routes your queries, your sector specialist reasons through complex analysis, and Claude fills knowledge gaps — all managed by a trust proxy that never leaks your data. One app, every AI, your rules.
  • ★ Use AI at work without violating company policyYour company's data stays on your machine — never uploaded, never shared, never stored on someone else's server. Get the productivity of AI while staying fully compliant with your organization's security policies.

Have questions? Need a demo? Want to talk about your use case?

We'd love to hear from you.

Reach out to us — [email protected]