LumaDock

29/09/2026

OpenAI's new always-on agents, called , launched this week. Personal ChatGPT Pro accounts in the UK, Switzerland and the EEA can't get one yet.

If you want an agent that keeps working after you log off, you can run your own. Hermes Agent is open source, runs on a server you choose and works with the model you choose. We compared it with dots, including the maintenance you take on when you self-host.

23/09/2026

Paying per API call to classify support tickets or score signups? does it on your own GPU server for a flat price, and existing Jev clients switch by changing one URL. We wrote up the install, the systemd setup and the three cases where it's less accurate than .

16/09/2026

A new AI model launched this week that can't write a single word. Jev, from TypeSafe AI, only answers questions with fixed options, and it answers them fast enough to play Doom in real time. It's also about 75 times cheaper per decision than a normal LLM on the same task.

We explain what it does, where it fits, where it doesn't, and which open-source tools do the same job on your own server.

$0.042 per million tokens, 70 to 500 ms, calibrated confidence, closed weights, a waitlist. And what runs on a VPS.

15/09/2026

is a free, open-source tool that tells you if ChatGPT and Gemini recommend your business or a competitor. It runs on a small VPS. We put it on Ubuntu 24.04, locked it behind a password (it doesn't come with one) and wrote every step down in English, because the official guide isn't.

14/09/2026

Running on your own server? A lot changed with 2.x and most tutorials never caught up: renamed Docker tags, $env blocked in Code nodes, MySQL dropped, and from October the npm install path goes away entirely.

Our self-hosting guide is written for the version that's actually out there, with a sizing table, a compose file with Postgres and Caddy, the env vars you need on day one and the rest of the stack (backups, queue mode, monitoring). We run n8n ourselves, so the numbers in it come from a live instance.

2 vCPU and 4 GB is the production floor. Pin n8nio/n8n:2.38.5 and set N8N_ENCRYPTION_KEY before the first start.

10/09/2026

Meta's new Muse Spark 1.3 is out. We looked at it from the angle we care about: what can you run on your own server?

Turns out the model everyone is talking about is API-only, but one of its smaller siblings runs on a 24 GB GPU with one command. There's also a pricing catch you'll want to know about before you paste a client's codebase into the cheap tier.

Spark 1.3 stays on Meta's API. Glimmer 30B is the Muse you can run yourself, on a 24 GB GPU with one Ollama command.

03/09/2026

New capacity is live in ! Our CEO was on site at our partner's firstcolo's Werkhaus facility, home of our availability zone, to see the expansion through in person. Plenty of CEOs would sign off on a purchase order and wait for the photos. Ours prefers to stand next to the racks 😊

Frankfurt has earned every euro we've put into it. The facility sits next to DE-CIX, the largest internet exchange in the world, which is why latency from here to most of central Europe is measured in single digits. It also runs on 100% renewable power and carries TÜV Saarland Level 3 certification.

For customers the takeaway is simple: more room in Frankfurt for the projects you've been planning.

https://lnkd.in/dmzbJ3fa

31/08/2026

Fancy running your own monitoring? on a is a genuinely satisfying weekend project, and we've put together the full walkthrough.

Install, first login, plug in , build a dashboard, get alerts landing in your inbox. It's written against Grafana 13.2, so everything in it is current.

The part we enjoyed writing most was the dashboards section. Six to eight panels you'll read beats forty you scroll past, and getting there takes less effort than people expect.

27/08/2026

The mystery model from last week has a name and a price tag, plus a spec sheet everyone keeps misreading. GLM-5.3-Flash costs $0.15 per million tokens and its benchmarks held up under independent testing. But "open weights" doesn't mean it runs on your gaming PC: the full checkpoint needs more VRAM than most companies own.

We wrote the developer's guide: which numbers are verified and what hardware you'd really need.

27/08/2026