Vehla does not ship with a model you rent from us. That is a deliberate choice, and the cost of it is that you have to make one decision before the AI actions do anything. Here are all six options, in the order most people should consider them.
Which route to pick
| IF YOU WANT | ROUTE | SETUP |
|---|---|---|
| Nothing leaves the Mac | Catalog model | One download |
| Best possible answers | Cloud key | Two minutes |
| Models you already pulled | Ollama | One URL |
| A specific quantization | MLX import or LM Studio | Longer |
| Vision or web research | Cloud key, required | Two minutes |
Nothing stops you from configuring several. Most people end up with one local model and one cloud key, and switch between them from the palette footer depending on the task.
1 · Catalog models
The default path, and the only one that works with the network off. Vehla ships an MLX inference runtime for Apple Silicon and a small catalog of Gemma 4 builds.
2. Download Gemma 4 E2B · 4-bit (~2.5 GB) to start. It is the Free-plan default.
3. Select Local AI from the palette footer.
4. Select text anywhere, press your palette hotkey, run summarize.
If you have the disk and the Premium tier, E4B at 4-bit (~4.5 GB) is a noticeable step up in writing quality, and the 12B Coder builds are worth it if most of your prompts are code. Free covers catalog models up to 2B parameters.
Two settings worth finding early: you can relocate model files to another volume if your internal disk is tight, and you can have models warm-start or offload when idle so they are not holding memory all day.
2 · Your own MLX model
Premium lets you import compatible MLX models rather than only using the catalog. This is the route if you have a specific fine-tune, a quantization you trust, or a model your team standardized on. It runs through the same on-device runtime, so the privacy behavior is identical to the catalog path — no network, no fallback.
The constraint is compatibility rather than size: it has to be an MLX-compatible build, and text-only. Import it from the same Local models screen and it appears in the footer picker next to the catalog entries.
3 · Cloud keys (BYOK)
Five providers are supported directly: OpenAI, Anthropic Claude, Google Gemini, DeepSeek, and OpenRouter. OpenRouter is the escape hatch — it accepts any model id it serves, so it covers models the other four do not.
2. Paste it into Settings → AI → Cloud API keys.
3. Use the test button. Do this — it catches billing and permission problems before they look like app bugs.
4. Pick a model in the provider's model picker, then select the provider from the palette footer.
Keys are stored in macOS Keychain and requests go from your Mac straight to the provider. Vehla is not in the path, which also means we cannot see your usage or help you with a rate limit. Providers you have not configured are greyed out in the footer rather than hidden, so you can see what is available.
A cloud key is required, not optional, for three things: vision actions, web-search agent actions through your Brave key, and anything where you want frontier-model reasoning.
4 · Ollama
If you already run Ollama, this is the fastest local route because the models are on disk already.
2. In Settings → AI → Ollama, point Vehla at the endpoint — usually http://localhost:11434.
3. Enter a pulled model tag, for example llama3.2:3b.
4. Select Ollama from the palette footer.
The one gotcha: Ollama has to be running before you select it. If the server is down the route fails rather than falling back, which is the correct behavior but does look like a bug the first time.
5 · LM Studio
Same shape as Ollama, different server. Start LM Studio's local server, copy its base URL into Vehla's LM Studio settings, choose a served model, and select it from the footer. Requests go to your local LM Studio server and nowhere else.
This is the route to use if you care about picking exact quantizations and inspecting how a model behaves — LM Studio's model management is better than anything we would build, so we did not build it.
6 · Lekh AI Pro
Lekh AI Pro is a companion app. With it installed, Vehla can discover its compatible local models, use Lekh as a local provider, hand off context with Continue in Lekh, share portable personas, and run Deep Research jobs that report progress, survive a relaunch, and deliver reports and source files back to Shelf.
Lekh routes stay hidden in the interface until the companion app advertises the required integration capability, so if you do not have it installed you will never see these options. Deep Research is the reason most people take this route.
Switching mid-task
The palette footer is the route switcher, and switching is designed to be a per-prompt decision rather than a settings trip. The practical pattern: local for a first pass on something sensitive, then switch to cloud for the same text if the local answer is not good enough and the content turns out to be harmless.
The footer always names the active route, and the context label above the input names what will be sent — selected text, clipboard, front app, attached files. Read both before pressing Return on anything you would not want to send out.
Passing a privacy review
If you are installing this on a work machine, the configuration that answers the most questions is: a catalog model downloaded, Local AI selected, no cloud keys configured at all, MCP servers left empty, and web-search agent actions unused. In that state text AI actions, Flow dictation, Read Aloud with Kokoro voices, clipboard OCR, Recall retrieval, and every native command run without a network. Licensing and update checks still use one.
With no cloud key configured, the vision actions and web-research actions cannot run. That is the tradeoff, stated plainly, and it is usually the right one for a work laptop. The full account of what is and is not local is in What actually stays on the device.
Troubleshooting
Full reference lives in the docs. If something here is out of date, email kailaconsulting@outlook.com and we will fix the post.
Common questions
Do I need an API key to use Vehla's AI?
No. Download a catalog model and text AI actions run entirely on-device with no key and no network. A cloud key is required only for vision actions, web-search agent actions, and frontier-model reasoning.
Which cloud providers does Vehla support?
OpenAI, Anthropic Claude, Google Gemini, DeepSeek, and OpenRouter — the last acting as an escape hatch since it accepts any model id it serves. Keys live in macOS Keychain and requests go straight from your Mac to the provider.
How do I connect Vehla to Ollama?
With Ollama running, set the endpoint in Settings → AI → Ollama, usually http://localhost:11434, enter a pulled model tag such as llama3.2:3b, and select Ollama from the palette footer. If the server is down the route fails rather than falling back.
What is the most private configuration for a work Mac?
A catalog model downloaded, Local AI selected, no cloud keys, no MCP servers, and no web-search actions. Everything except licensing and update checks then runs without a network. The tradeoff is that vision and web-research actions cannot run at all.