Matrix LLM · iPhone, iPad, Mac and Windows
Ask your documents on your Mac, PC or iPhone, offline.
Add contracts, regulations or reports and ask in plain words. Every answer shows the exact passage it is based on. The models run on the device, so your documents stay there.
Free to download for iPhone, iPad, Mac and Windows. The interface is in 22 languages.

The assistant
You can ask it to operate the app.
You don't need to know what a vector or a knowledge graph is. Ask in plain words: "Create a collection called Contracts", "Check all supplier contracts for notice periods", "Switch to the Paper style". The assistant uses the same tools an AI agent gets and shows every step.
Asks before it acts
Anything that changes something first appears as a card: what it will do and how much work it is, for example "About 40 model requests". It runs only after you tap Run.
Cannot delete
Deleting is not one of its tools. The settings it may change come from a fixed list, never keys or endpoints.
Shows the result
After a step it opens the result: the collection, the table run, the term in the graph, the document.
Built-in help
"How do I add files?" is answered by any model from a built-in guide, with a "Show me" link to the right place.
Stops after two failures
If the same action fails twice, it stops and says what went wrong and what you can do.
The limit
Working the app needs a model that can call tools. On Apple devices, Apple's on-device model answers the help questions, not the app tasks.
The first five minutes
From download to the first answer with sources.
Pick a model
On Apple devices that offer it, Apple's model needs no download. Otherwise, and on Windows, one tap loads a local model. The search model downloads in the background; after that everything works offline. On Windows, the search model and, if you want one, a language model come from Hugging Face once, after you agree.
Try a sample collection
Three ready-made sets: a German sample contract, the official accident file of the US Airways flight that landed on the Hudson, and Wikipedia articles on the Panama Papers.
Add your own files
PDF, Word, Excel, PowerPoint, e-mails, e-books, ZIP archives, photos and scans (with text recognition: Apple's built-in on Apple devices, a model that downloads on first use on Windows). On Apple devices you can also share a file to Matrix from any other app; on Windows, drop files onto the window.
Ask about a file right away
Attach it to a question with the paper clip. Afterwards keep it as a collection or discard it.
Tap a citation
It opens the exact passage; in PDFs, on its page.
The principle
The model reads. Everything that can be checked is plain code.
The language model is used only where language has to be understood. Splitting documents, following references, verifying quotes, correcting sources and logging every step is done by ordinary code. It is reproducible, and no AI judgement is involved.
On this page, AI means the model decides,rules means deterministic code does.
Every answer shows its sources
Down to the verbatim quote in the original. You can click it and check it.
Your data stays local
It runs with local models on the device or inside your own network. Nothing has to leave.
Everything is logged
Every search, every hit and every automatic correction is logged and can be replayed.
Ask a question
It searches before it answers.
You ask once, and the model searches the documents itself. Plain code sets its limits and checks the result.
Prepare the documents rules
They are split along their real structure: articles, sections. Each piece gets an embedding, a vector that stands for its meaning, so search works on meaning and not only on words. References such as "pursuant to Art. 92(1)" are linked to their target.
Search, read, search again AI
The model writes its own search queries, opens passages, follows references and searches again with the terms it found, until it has enough evidence or its budget is used up.
Guard rails rules
A search budget caps the cost. Watchdogs stop it from going in circles. An answer given without any search is thrown away.
Answer only from what it found AI
Every statement is tagged with its source number: [1], [2], [3].
Check the answer rules
The checks in the next section run on every answer. Then everything is shown with clickable sources and a full log.
Citation checks
It checks every citation.
After every answer, a series of plain-code checks goes over each citation.
Does the source exist?
Every [n] is checked against the list of sources that were actually retrieved.
Is the quote really there?
Quoted words are searched for in the original. If they are not found, the answer gets a warning sign.
Does that source say it?
Two independent signals: shared key terms (legal references count three times) and closeness in meaning.
Fix only when both agree
Only then is the marker moved to the right source and labelled "→ 3". Otherwise there is just a hint. The answer text itself is never rewritten.
Tables
The same question for every row, as a table.
Pick a collection or upload your own list and ask the same question for every row. Each answer has to fit the columns you define: text, number, date or choice.
Answers are validated
Missing or wrongly typed fields are shown in the cell.
Lists become rows
Five items in one answer become five rows, each still linked to its source passage.
Redo single rows
Run single rows again with a stronger model. Duplicates are merged. Export to Excel.
Typical uses: a list of every obligation in a regulation with addressee, deadline and evidence · the same check across hundreds of contracts · a first screening of a customer list.
The knowledge graph
Who relates to what, with a quote for every link.
The model reads every passage and extracts things and their relations: who is responsible for what, what mitigates which risk, who owns whom. Every link carries a verbatim quote from the original.
Quotes are verified
A link whose quote can't be found is repaired first, not silently dropped. If the quote says the opposite ("must not…"), the link is flagged.
Duplicates merged
"BaFin" and "Federal Financial Supervisory Authority" become one node. Every merge is logged and can be undone by hand.
Your own vocabulary
The raw labels stay as they are. A switchable "lens" maps them onto your own vocabulary, without extracting again.
Topics and summaries
The graph is split into clusters, each with a short summary, which works as a table of contents for the whole collection. Export to PowerPoint.
Failures are retried
Passages that fail go to quarantine and are retried. Every finding goes to a review queue.
Quality depends on the model
A local 30B-class model gives a fast first map, good for orientation but not proof. A frontier model gets close to "audit-ready".
Platforms
One core for Mac, Windows, iPhone and browser.
Everything important is in a single core written in Rust. The apps on top are thin.
iPhone and iPad
Fully offline: search runs on the Neural Engine, chat runs with a local model on the GPU. Files come in through the share menu of any app.
Mac
The same app. In addition, Claude Code and other AI agents can use the running app through MCP (Model Context Protocol, the standard way for agents to call tools).
Windows
The same app for 64-bit PCs. Local models run on the graphics card through Vulkan (AMD, NVIDIA, Intel), otherwise on the CPU. The MCP access for agents is there as well, off by default. It needs 8 GB of memory, 16 GB are recommended.
Browser
A Linux container in your own network. Users install nothing.
Any model
Use any model, or only local ones.
All model connections are managed in one place, and each is assigned to the task it does well.
On the device
Models from Hugging Face, searched and loaded inside the app, on every platform. On Apple devices also Apple's built-in model, without a download.
Search models
Embedding and re-ranking models for search, also offline.
Company gateway
OpenAI-compatible or Azure endpoints behind your own infrastructure.
Cloud providers
Direct providers where that is allowed, chosen deliberately, case by case.
Limits per model
Requests per minute, token budgets, parallel requests, so gateways and budgets are not overloaded.
Only what the model supports
Tool use and reasoning are switched on only for models that support them.
When things fail
Tested against real outages.
Load balancing
Several endpoints work as one pool. Throttled ones are routed around before a fallback model takes over.
Streaming on slow gateways
When a proxy gets slow, the app switches to streaming so the connection doesn't time out.
Retries
Temporary errors are retried with growing waits. Cases that keep failing are skipped instead of stopping the run.
Fallback model
If a model returns unusable output, a backup model takes over, only for that one passage.
Background jobs
Long runs continue in the background with visible progress, and the app stays usable.
Temperature
On iPhone and Mac, local models watch the device's temperature and slow down instead of overheating.
By the numbers
What I measured.
I track search quality with fixed sets of test questions and the passages each one should find. Quality is judged on those, not by impression.
Also in Claude Code
The same engine as a Claude Code plugin.
Claude Code can use Matrix as a tool. On 442 pages of EU regulations it answered in 42 instead of 126 seconds per question on average, with less than half the input tokens.
About the pluginWhere it stands
Free on the Apple App Store and in the Microsoft Store.
For iPhone, iPad, Mac and Windows. Asking and tables work with the model on the device. For the knowledge graph, a local model gives a quick first map and a stronger model a more reliable one. The checks run either way and show where the output can be trusted.
On Windows 10 (version 1809) or later: winget install 9PGLD9C1DQJW.
What the app stores and when anything leaves your device: privacy policy.