Yes. Vox is 100% free and open source under the MIT license. No subscription, no usage limits, no hidden costs.

Does Vox work offline?

Yes. After the initial model download, Vox runs entirely on your Mac with no internet connection required.

Vox runs locally on your hardware. Your conversations, files, and data never leave your device. No cloud, no telemetry.

What can Vox do that ChatGPT cannot?

Vox controls your desktop — it manages files, sends emails, replies to iMessages, captures screen regions, and runs background agents. It works offline and keeps all data on your device.

Does Vox work on Windows or Linux?

Vox is Mac-first today. The core architecture is platform-agnostic, and Windows and Linux support is the top contributor priority.

Vox vs ChatGPT for Students — Why One Pass Isn’t Enough

Every student has tried it: paste your lecture notes into ChatGPT and ask for a study guide. It works great — for short inputs. But the moment you throw in a full semester's worth of material, a 300-page textbook, or a stack of research papers, the cracks show. Here's why, and what Vox does differently.

The single-pass problem

ChatGPT, Claude, Gemini, Copilot — they all process your input in a single forward pass. You paste text in, the model reads it once, and generates a response. This architecture has three consequences for students:

Context window limits. Even models with 128K token windows can't hold a 300-page textbook. You have to manually split content into multiple conversations, losing continuity between them.
Attention degradation. Research consistently shows that LLMs pay less attention to content in the middle of long inputs. Your Chapter 8 notes get less coverage than Chapters 1 and 15.
No verification. The model generates once and stops. It never goes back to check whether it missed a concept, skipped a formula, or hallucinated a definition.

How Vox handles the same material

Vox spawns a background worker — an autonomous agent that iterates through your material in multiple passes with a structured workflow:

ChatGPT

Single forward pass through input
Manual copy-paste into chat window
You split large docs yourself
No memory between conversations
Response generated once, no review
Requires you to sit and prompt

Vox

Multi-pass iterative processing
Reads files directly from your desktop
Auto-chunks docs (1,400 chars, 200 overlap)
Persistent journal tracks all discoveries
Verify phase checks against source
Background worker — you walk away

A real example: finals week

“Computer, I have finals in three days. Process my entire Organic Chemistry folder — all lecture notes, textbook PDFs, and lab reports — and create a comprehensive study guide organized by topic.”

Vox — Background Agent: Organic Chemistry Finals

✓Scanning folder: 23 files (4 PDFs, 12 markdown notes, 7 lab reports)Done

✓Indexing all documents into knowledge baseDone

✓Planning: identified 14 topic areas from file analysisDone

✓Pass 1: Extracting core concepts (functional groups, reactions, mechanisms)Done

✓Pass 2: Cross-referencing lab reports with lecture theoryDone

✓Pass 3: Identifying gaps — 3 topics mentioned in notes but not fully coveredDone

✓Pass 4: Deep-dive on gap topics using textbook sourceDone

✓Verification: all 14 topics confirmed covered, generating outputDone

Vox

Your Organic Chemistry study guide is ready. 52 pages organized by topic: • 14 topic sections with textbook + lecture note cross-references • Reaction mechanism flowcharts for all 23 named reactions • Lab practical summaries tied to theoretical concepts • 85 practice problems (your professor's style, based on past assignments) • 3 topic areas where your notes were thin — I supplemented from the textbook Saved to your Desktop. I also found 3 contradictions between your notes and the textbook — flagged in yellow.

Notice what happened: the worker found gaps in your notes and went back to the textbook to fill them. It cross-referenced lab reports with lecture theory. It flagged contradictions. None of that happens in a single pass.

The journal makes the difference

Vox's background worker maintains a persistent journal throughout the task — a structured working memory that tracks:

Understanding: What the agent has learned about your material so far
Discoveries: Key concepts, formulas, and relationships found during processing
Completed steps: Which sections have been fully processed
Current plan: What to process next and why
Blockers: Any issues or missing information

When the agent processes Chapter 12, it has full awareness of what it found in Chapters 1-11. When ChatGPT processes Chapter 12, it's fighting attention degradation from a massive context window.

What about uploading PDFs to ChatGPT?

ChatGPT Plus lets you upload files, yes. But it still processes them in a single conversation turn. Upload a 300-page PDF and ask for a study guide — you'll get a broad overview, not detailed exam-ready material. The model doesn't iterate. It doesn't go back and verify. It doesn't fill gaps.

Vox's indexer processes that same PDF differently: splitting it into overlapping 1,400-character chunks with smart paragraph-boundary detection, running four concurrent workers, and storing every chunk for precise retrieval later. The background agent then works through the indexed material systematically, not all at once.

Note

Vox also processes files locally on your machine. Your study materials never leave your computer during indexing — important if you're working with proprietary course materials or unpublished research.

When to use what

ChatGPT is fine for quick questions — "explain this concept" or "rephrase this paragraph." For anything involving large volumes of source material and thorough output, the multi-pass approach produces measurably better results: