Point Junyr at your own LLM server (Ollama, LM Studio, vLLM…) and every realtime AI feature runs on your hardware. AI works even on your most confidential inboxes — without ever leaving your machine.

Bring your own AI: run Junyr's AI on your own server

Junyr lets you point every realtime AI feature at your own LLM server. When it's reachable, summaries, search, smart replies and agent chat run on your hardware instead of the cloud — and your data never leaves your machine. It's per-user, and it's available on every plan.

---

The sovereignty problem with "AI for business"

Most AI products ask you to make a trade: useful automation, or control over your data — pick one. For an inbox full of salaries, IBANs, contracts and client secrets, that's not a real choice. So we built the other option.

Under Settings → Local LLM, you give Junyr the address of a model server you control. From that moment, every eligible AI request runs there.

---

Any server, your choice

Junyr speaks the OpenAI-compatible API, so it works with whatever you already run:

• Ollama, LM Studio, vLLM, or any OpenAI-compatible gateway
• On your LAN, your Tailscale network, or your own datacenter

A built-in connection test confirms reachability and discovers the models your server exposes. Per-modality capability probes run a real inference for each capability you care about — chat (text→text), vision (image→text), transcription (speech→text) and embeddings — and report whether each one works, with a sample and a latency reading.

---

What runs locally

When your server is online, Junyr routes essentially all of its realtime AI to it:

• Compose review and reformatting before you send
• AI mailbox search ("find the renewal quote from spring")
• Smart actions on an open email
• Agent chat and project discovery
• Workflows — and even the nightly Reflexions pipeline

There's no separate "local mode" to babysit. It's the same product; the inference just happens on your box.

---

True sovereignty, not marketing sovereignty

This is the part that matters. Junyr has three per-inbox confidentiality tiers — Totale (no cloud AI at all), Sécurisée (AI with personal data masked first), and Simple. A reachable local server changes what's possible:

• It lifts the Totale block. Your most confidential mailboxes get summaries, smart replies and search — because the request physically never leaves your hardware.
• It skips the Sécurisée masking step for calls that run locally. Salaries, IBANs and contracts stay intact, because there's no third party to hide them from.

And when your private server is offline, Junyr fails safe: confidential-tier calls return an error rather than quietly falling back to the cloud. Nothing leaks because the system would rather do nothing than betray the tier you chose.

---

No hardware? Bring your own cloud key instead

Running a model server isn't for everyone. If you'd rather use a commercial model but still own the relationship, plug in your own Gemini, Claude or Mistral API key (encrypted at rest). You control the provider and the bill.

One honest caveat: your own cloud key is still cloud. It's gated exactly like Junyr's platform cloud — Totale blocks it, Sécurisée masks personal data first. Only the local server is fully sovereign. We don't blur that line.

---

Private by design

• SSRF-guarded endpoint validation blocks cloud-metadata addresses while still allowing LAN and Tailscale hosts
• Your bearer token is write-only and stored encrypted — it's never sent back to the browser
• Everything is per-user: your colleague's choice of provider is independent of yours

---

Available on every plan

This isn't an "Enterprise" upsell. Sovereign AI ships on Solo, Studio and Scale alike. Hosted in Europe by default, on your own hardware when you want it — that's the whole point.

Set it up in Settings → Local LLM and run your business on AI you actually control.