PrismOS-AI: a local-first desktop AI that reads your files and keeps the answers on disk
Desktop AI that runs on your laptop. Drop a file, ask a question, it remembers. Nothing leaves the machine.
At a glance
- What is it?
- PrismOS-AI is a Tauri 2 desktop app that answers questions about your own documents using a local Ollama model. The README is unusually specific about what leaves the machine, and unusually honest about what the installers cost you.
- Who is it for?
- PrismOS-AI suits engineers who already run Ollama and want document Q&A without sending files to a hosted API, and who are willing to clear a quarantine flag or build from source. Skip it if you need signed installers, an ARM Linux build, or a hosted model's quality ceiling.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 27 days ago.
- What is it written in?
- Mainly Rust, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem PrismOS-AI targets, and who feels it
Most document question-answering tools assume you will upload the document somewhere. PrismOS-AI takes the opposite position: you drop a file, ask a question, and the answer comes from an Ollama model running on the same laptop. The README frames the whole product in one line, "Open the lid, ask, close the lid. Your AI runs on your laptop, zero bytes leave the machine."
The audience is narrower than "everyone who wants an AI assistant." It is people who already have a reason to keep files local: internal design docs, contracts, incident notes, anything under a policy that forbids pasting into a hosted chat window. The README also targets a second group indirectly: engineers who want to evaluate the idea without installing a GUI, which is why a standalone CLI binary exists alongside the desktop app. If neither of those describes you, the friction below is not worth paying.
Architecture: Tauri shell, Rust core, SQLite knowledge graph, Ollama on 11434
The stack is Tauri 2 plus React 18 plus Rust, with TypeScript on the front end and a Vite build. That split matters for the privacy claim, because the webview is where a network call would have to originate, and Tauri lets the app constrain it declaratively.
The README's strongest claim is not a promise but a configuration file. The Content Security Policy in src-tauri/tauri.conf.json lists exactly one external origin in connect-src:
connect-src 'self' http://localhost:11434 http://127.0.0.1:11434Port 11434 is the local Ollama daemon. Because no other host appears in the allow-list, the UI cannot reach one, and the enforcement happens in the OS webview rather than in application code. That is a meaningfully different guarantee from a README sentence saying the app is offline.
On disk, the app keeps what it learns in a knowledge graph backed by SQLite. The front end pulls in react-force-graph-2d, which suggests the graph is rendered visually, and react-markdown for answer formatting. Retrieval happens locally against that store, so a later question can draw on a document indexed in an earlier session without re-reading the file.
Installing PrismOS-AI, and a first question through the CLI
The README is upfront that the installers are not code-signed. On macOS you will see "PrismOS-AI is damaged and can't be opened" or a malware warning, and the documented fix is to clear the quarantine attribute:
xattr -rd com.apple.quarantine /Applications/PrismOS-AI.appOn Sequoia and later, the README notes that Control-click then Open no longer works, so this command or System Settings then Privacy & Security then Open Anyway is the path. On Windows, SmartScreen shows "Windows protected your PC" and you click More info then Run anyway.
The shell installer resolves the correct asset for your architecture, verifies its SHA-256 against the digest GitHub publishes, and aborts on mismatch. It also bootstraps Ollama with qwen3:4b if you do not already have it:
curl -fsSL https://raw.githubusercontent.com/mkbhardwas12/prismos-ai/main/scripts/install.sh | shExpect roughly 15 MB for the app, the Ollama runtime, and about 2.5 GB of model weights. The README budgets 5 to 15 minutes on a decent connection and around 4 GB of free RAM to run qwen3:4b comfortably.
If you would rather not install a desktop app to evaluate the idea, the CLI avoids the GUI and the Gatekeeper step entirely. Build it, check that the daemon is reachable, then ask a question:
cargo build --release --bin prismos-cli
./target/release/prismos-cli health
./target/release/prismos-cli models
./target/release/prismos-cli ask "explain WASM fuel metering in one paragraph"The health subcommand reports whether the daemon is up, models lists what is pulled locally, and ask sends a prompt. It is pipeable, so a file can be fed through stdin with the --stdin flag and a --model override. PRISMOS_MODEL and PRISMOS_OLLAMA_URL change the defaults without editing anything.
Where the offline claim has edges
The README does not claim the app never touches a network, and it names three exceptions. The installer downloads from GitHub, and Ollama downloads model weights the first time. After that, the README states, no network is required or used.
The other two are optional features that are off by default. Email Keeper connects to your own IMAP server if you configure it. Web Research fetches pages, but only URLs you type into chat, over HTTPS, with localhost and LAN addresses refused. It is double-gated: a Settings toggle plus a Rust-side gate that hard-refuses fetches while disabled. There is no search engine involved, and the README says nothing is sent in the background. Saying "explore" also follows the most relevant links found on the pages you named, bounded, with every followed link passing the same gates. What gets read is indexed into the local SQLite graph so later answers can retrieve it.
The zero-network alternative to Web Research is built in: open the page and ask PrismOS to read your screen, using local vision only. If you leave both optional features off, the README's position is that the app has no reason to open a socket. That is a design argument, not a proof, and the CSP is the part you can actually inspect.
Unsigned installers, missing ARM Linux builds, and the wrong-tool cases
The most concrete limitation is code signing. The README states plainly that the certificates cost more than the project currently justifies, and that signing is on the roadmap. That is an honest trade, but it has consequences: on managed corporate laptops, Gatekeeper and SmartScreen overrides are often blocked by policy, which makes the desktop app unusable regardless of how good the local inference is. Building from source is the documented escape hatch, and the README says the build is reproducible from a clean checkout.
Platform coverage has a gap. The Releases table lists Windows x64, macOS Apple Silicon, macOS Intel, and Linux x64. Linux ARM is explicitly not published, with build-from-source as the only route. If you are targeting ARM servers or a Raspberry Pi, you are compiling it yourself.
There is also a resource floor that the README states rather than hides: around 4 GB of free RAM for qwen3:4b, plus 2.5 GB of weights on disk. On a laptop with 8 GB total and a browser open, that is a real constraint. And the model itself is a 4B parameter model at Q4_K_M. For summarization and retrieval over your own notes it is workable; for tasks that need long-context reasoning or precise code generation, a hosted frontier model will beat it, and no amount of local-first architecture changes that.
How PrismOS-AI differs from a hosted RAG stack
The obvious alternative is a hosted retrieval-augmented setup: embeddings and a chat model behind an API, with a vector store you manage. The difference is not just where inference runs, it is where the trust boundary sits.
With a hosted stack, the document leaves your machine at ingestion time and again at query time. You get a stronger model, managed scaling, and no 2.5 GB download. You also get a data processing agreement to read and a vendor to depend on. With PrismOS-AI, the document stays on disk, retrieval runs against a local SQLite knowledge graph, and the model runs in Ollama on port 11434. You trade model quality and operational convenience for the ability to pull the Ethernet cable and keep working, which the README invites you to actually try.
A second comparison point is a plain Ollama setup with a chat UI. That gets you local inference but not persistence: each session starts cold, and nothing is indexed for later retrieval. PrismOS-AI's knowledge graph is the part that makes a second question cheaper than the first. If you never ask follow-up questions across sessions, the graph adds storage and complexity you will not use.
Maintenance, licence, and what upgrading costs
The repository is not archived, and the last push was on 2026-08-20. The release history shows v0.5.1 as the public launch on 2026-03-04, v0.5.2 the same day, and v0.6.0 on 2026-08-12, so the cadence is irregular rather than steady: a cluster of releases in March, then a gap of roughly five months before 0.6.0. Plan for a project that moves in bursts.
The licence is MIT, which is permissive and places few obligations on how you redistribute or modify the code. The repository also carries a NOTICE file, which is worth reading alongside LICENSE if you plan to vendor any part of it. Nothing here is legal advice; if the app ends up in a commercial distribution, have someone check the NOTICE and the licences of the bundled dependencies.
The upgrade path has two moving parts. The front end includes @tauri-apps/plugin-updater, so the app can check for its own updates, and the CHANGELOG.md at the repository root is where version-to-version changes are recorded. Separately, Ollama and the model weights update on their own schedule, so a PrismOS-AI version bump does not tell you whether qwen3:4b changed underneath you. The README does not document rollback, so if a new release misbehaves, the documented recovery is reinstalling a previous asset from the Releases page.
Editorial conclusion
PrismOS-AI suits engineers who already run Ollama and want document Q&A without sending files to a hosted API, and who are willing to clear a quarantine flag or build from source. Skip it if you need signed installers, an ARM Linux build, or a hosted model's quality ceiling. Before adopting, check the CSP allow-list in src-tauri/tauri.conf.json, confirm you have roughly 4 GB of free RAM for qwen3:4b, and decide whether the Email Keeper and Web Research features will stay off, since they are the only two paths off localhost.
Frequently asked questions
Does PrismOS-AI send any of my data to the internet?
The README states that no bytes leave the machine once the app and model weights are installed, and points to the CSP in src-tauri/tauri.conf.json, whose connect-src allow-list contains only localhost:11434 and 127.0.0.1:11434. Two optional features can reach beyond localhost: Email Keeper connects to your IMAP server, and Web Research fetches URLs you type. Both are off by default.
What model does PrismOS-AI use, and how much RAM does it need?
The install script bootstraps Ollama with qwen3:4b if you do not already have it, and the README puts the weights at about 2.5 GB. It recommends roughly 4 GB of free RAM to run that model comfortably. The CLI accepts a --model flag, and PRISMOS_MODEL overrides the default.
Can I use PrismOS-AI without installing the desktop app?
Yes. The README documents a standalone binary built with cargo build --release --bin prismos-cli that talks straight to your local Ollama daemon, with no GUI and no Gatekeeper quarantine step. It exposes health, models and ask subcommands and is pipeable through stdin.
Community notes