Proton’s Lumo Desktop tops a week of AI on your terms, but every connector starts on auto-approve

20-11-2026

Proton released Lumo Desktop on 17 November, with connectors that read files and run shell commands, all set to auto-approve by default. In the same week Hugging Face moved chat models to OpenRouter, llama.cpp reached 1.0 and HeyGen launched Avatar 6.

Written by:

Jorick van Weelie

Marketing Lead at DataNorth | AI Enthusiast & Tech Storyteller

proton lumo desktop and three more ai releases of week 47
Sign up for our Newsletter

Proton’s AI assistant Lumo arrived on Windows, macOS and Linux on 17 November, with connectors that read local files, run shell commands and work in Google Drive and Jira. It tops a week in which Hugging Face moved its language model traffic to OpenRouter, llama.cpp reached version 1.0.0 and HeyGen released Avatar 6. Three of the four give you more say over where AI runs, and two of them come with a setting your team should check first.

Lumo Desktop can run shell commands without asking until you change one setting

Lumo Desktop puts Proton’s assistant in an app with connectors. They cover local files, the shell, Proton Mail through Proton Mail Bridge, Google Drive, Jira, Confluence and any MCP server you add. MCP is the open standard that lets an assistant use other tools. Scheduled tasks run a request at a set time, and Skills store instructions Lumo reuses.

Proton’s support page says every connector action is set to Auto-approve by default, and connectors are on in every conversation. For each action you can pick Auto-approve, Ask human or Deny. An assistant that reads web pages and may run commands without asking is exactly what prompt injection targets: hidden instructions in a page or file that the model then follows.

Proton says it never trains models on your conversations or connected data, that chat history has zero-access encryption, and that Lumo runs on European infrastructure under Swiss privacy law. The source code is public, but Proton names no licence, no prices and no models behind Lumo.

This is the release for organisations that want an assistant outside the big US clouds and are willing to set its permissions first. You download it from Proton’s website and sign in with a Proton account. Before anyone links a work account, set shell and file-writing actions to Ask human or Deny in the Lumo connector settings.

Hugging Face now routes chat models through OpenRouter and drops your own provider keys

Since 16 November, every language model request on Hugging Face Inference Providers goes through OpenRouter. Its router picks a provider per request on live price, speed and reliability, and retries elsewhere if the first one fails. Image, video and audio generation work as before.

Your code and token keep working, but Hugging Face flags one breaking change: you can no longer use your own API key for a provider with language models. You pay the provider’s price on your Hugging Face bill, with no markup. PRO users get $2 of credits a month, Team and Enterprise accounts $2 per seat. OpenRouter says 160 models from 38 labs are available this way.

Neither company says anything about data retention or where requests are processed, and your prompts now pass through OpenRouter as well as the provider. If your team has rules about who may see its data, switch off unchecked providers in your organisation settings. You can also pin one provider by adding its name to the model ID, such as :novita.

llama.cpp 1.0 promises a new major version for every breaking change

llama.cpp is the open-source engine that runs language models on laptops, phones and servers without a cloud service. On 13 November its team released the libllama and libggml libraries as version 1.0.0, three and a half years after the project began.

From now on, a change that breaks the public C interface gets a new major version, while features, hardware backends and model support arrive in minor versions. The project supports more than 150 model architecture variants. Neurohive, citing the GitHub release notes, reports that 1.0.0 removes 35 deprecated functions.

llama.cpp is MIT-licensed and ships through GitHub, Docker, Homebrew, Winget and Debian, among other channels. For the local and sovereign AI systems DataNorth builds, that version number matters: you can pin the engine and plan upgrades like any other infrastructure.

HeyGen Avatar 6 keeps one avatar talking for 28 minutes without a cut

Avatar 6 turns a single photo into a talking avatar video, optionally with a short recording of the person as a reference. Every frame is generated at 2880 by 1620 pixels, and HeyGen’s examples run 25 to 28 minutes from one audio track. You can add gestures by writing them next to the words, or change single words in an existing video in the speaker’s own voice.

In HeyGen’s own side-by-side test, raters picked the avatar as the real recording 42.2% of the time, where 50% would mean they could not tell the two apart. The test is HeyGen’s, and its report does not define all of its rating scales.

Avatar 6 has been in HeyGen’s app and API since 17 November. The first video is free, and paid plans start at $29 a month for 600 credits. HeyGen does not say how many credits a video costs, and its launch pages list no languages, Dutch included.

HR, learning and communication teams that make training videos should take a look. An avatar built from someone’s footage only renders after that person records an on-camera consent statement with a one-time code. Agree internally whose likeness may be used, and when a video must say it uses an avatar, before staff make digital versions of colleagues.

Try Lumo on documents you know well, with shell access set to Deny

Of these four, we would test Lumo Desktop first. For organisations that want to keep AI work away from US cloud providers, it is the most concrete option this week. Proton names no models, though, so answer quality is still an open question. Give it a folder of documents you know well, set the shell to Deny and every sending or deleting action to Ask human, and compare its answers with what you know. The other three can wait for a concrete need: a local model project for llama.cpp, a provider policy for Hugging Face and a likeness policy for HeyGen.

For more information, visit the official announcement of Lumo Desktop on the Proton blog.

Add DataNorth AI to your Google favorites