For IT departments
What Retriever downloads, and from where.
Motekai Retriever runs on the PC. Its search models come inside the installer. For answers written on the PC, the default, the first setup downloads an AI engine and a language model. This page lists each download, the hosts it comes from, its size and how it is checked, so you can decide whether to allow it.
Last checked against the app: 15 October 2026 · version 0.3.33
In short
- Indexing, search and answers run on the PC. The app, and the engine it starts, listen only on 127.0.0.1.
- The installer is about 840 MB. It carries the search models, so search needs no download.
- For answers written on the PC, the default, setup downloads two things once, over HTTPS: the engine from GitHub and a language model from Ollama's registry. If the user chooses a cloud AI provider at setup, neither the engine nor the language model is downloaded.
- On a PC without Microsoft Edge WebView2 Runtime, the app first runs Microsoft's installer for it, which downloads the runtime from Microsoft.
- After setup, answers work offline.
- After that, on the default local setup, the app contacts only our release bucket, for updates and revoked licence keys. It sends no documents and no questions, and there is no account and no telemetry.
What the installer carries
The search models ship inside the installer, so the app searches with no download at all: nomic-embed-text-v1.5, which turns documents and questions into something the app can search, and bge-reranker-base, which puts the best-matching passages first. Each file is pinned by its SHA-256 and size, and a build whose files don't match is refused. The installer is about 840 MB to download.
Versions before 0.3.33 downloaded these models during setup, from huggingface.co and Ollama's registry. Version 0.3.33 doesn't, and contacts huggingface.co for nothing.
What setup downloads
Only for answers written on the PC, the default. In this order. Each is downloaded once, for the Windows user who sets the app up.
1. The answer engine
Ollama 0.9.6 for Windows (ollama-windows-amd64.zip), the open-source engine that runs the language model on the PC.
| From | github.com, Ollama's release page. GitHub hands the file itself to release-assets.githubusercontent.com. |
|---|---|
| Size | 1.33 GB (1,330,113,375 bytes) to download |
| Checked by | Its SHA-256, pinned in the app: dff2167a87813899a761704a4240e5edf5766df7d6a5f7d4f27b7d36c6ea2f2f, the value Ollama publishes for this release. A file that doesn't match is deleted. |
2. The language model
qwen2.5:7b on a PC with more than 8 GB of memory, or qwen2.5:1.5b on one with 8 GB or less. It writes the answers. If the app can't read the PC's memory it takes qwen2.5:7b.
| From | Ollama's model registry, registry.ollama.ai, downloaded by the engine above. The registry sends the model files from a separate download host (see below). |
|---|---|
| Size | As listed on Ollama's page for qwen2.5. |
| Checked by | Ollama, which checks every file against the digest in the registry's manifest. The model is asked for by name, so it is whatever the registry publishes under that name. |
Microsoft Edge WebView2 Runtime
The app draws its window with Microsoft Edge WebView2 Runtime, a part of Windows that current Windows 10 and 11 PCs usually already have. On a PC without it, the app runs Microsoft's own WebView2 installer, which comes inside ours, before its window opens. That installer downloads the runtime from Microsoft's Edge update and download service. Our code hands it nothing of the user's: no documents, no questions, no settings.
Which hosts Microsoft's installer contacts is Microsoft's to say. Microsoft lists them for the Edge update service, which also installs and updates WebView2 Runtime, in its allow list for Microsoft Edge endpoints, including msedge.api.cdp.microsoft.com (updates) and *.dl.delivery.mp.microsoft.com (downloads). We haven't captured the installer's own traffic, so treat Microsoft's list as the authority. Its size is Microsoft's too. If you would rather install the runtime yourself, Microsoft offers it at developer.microsoft.com/microsoft-edge/webview2; the app finds it and skips this step.
Hosts to allow
All on port 443 (HTTPS), except Microsoft's: its list above gives its download hosts over both HTTP and HTTPS.
| Host | For |
|---|---|
| github.com | The answer engineSetup, for answers written on the PC |
| release-assets.githubusercontent.com | The answer engine (GitHub's file host)Setup, for answers written on the PC |
| registry.ollama.ai | The language modelSetup, for answers written on the PC |
| Microsoft's Edge update and download hosts | Microsoft Edge WebView2 Runtime, installed by Microsoft's own installerFirst launch, only on a PC without the runtime |
| pub-5f9fecebe96a48b48ec442fc9ffc1d27.r2.dev | Update note, revocation list, update packagesEach time the app opens |
After setup
Answers work offline. The engine, the models and the index are on the PC, and the app answers from them without a connection.
When it is online, each time the app opens (and no more than once an hour) it fetches two small public files from our release bucket, pub-5f9fecebe96a48b48ec442fc9ffc1d27.r2.dev: a note of the newest version, and the list of revoked licence keys. If there is a newer version, it downloads the update from the same bucket, checks its checksum and its Windows code signature, and installs it the next time the app opens. Those requests carry no documents, no questions, no licence key and no machine ID. If they fail, the app carries on.
Cloud AI is a choice made at setup, and local is the default. If the user connects their own cloud AI account, each question and the passages it matched go to that provider. Documents are still indexed and searched on the PC. The privacy notice has the detail.
Proxies
The app has no proxy settings of its own. It uses the proxy named in the Windows user's environment: set HTTPS_PROXY (for example http://proxy.example.com:8080) for the user who runs it, then open the app. The app's own downloads read it, and so does the engine it starts, for the model downloads. Ollama's documentation advises against setting HTTP_PROXY. Microsoft's WebView2 installer is Microsoft's: how it finds a proxy is up to Microsoft.
- Proxies that ask for a sign-in aren't supported. The app can't answer a proxy's request to authenticate (HTTP 407).
- Proxies that decrypt and inspect HTTPS work once their root certificate is in the Windows certificate store. The app and the engine it starts check certificates against that store, as browsers do, so a proxy root your organisation has already installed for them is trusted. We have tested this with a stand-in certificate authority, not behind a real inspecting proxy.
If the downloads are blocked
Search never waits on a download. If the engine or the language model is blocked, setup stops at that step, says which host it was downloading from, and offers Retry, Copy details for IT and What your IT team needs, which opens this page in the user's browser. Once the hosts above are allowed, Retry carries on from where it stopped.
It also offers Continue without written answers: setup goes on to the folder and the index, and the app shows the passages that match each question instead of writing an answer. Its home screen then offers Try the download again, which downloads the language model only when the user chooses it.
If WebView2 Runtime is missing and can't be installed, the app says so in a message that names Microsoft's download page and this one, rather than opening an empty window.
Rolling it out on many PCs, or on a network with no internet access? Email [email protected] before you start, and we'll talk it through. For what the app is and what it needs to run, see the system requirements.