We build it. You own it. Then we leave.
OddBox builds AI-managed applications shaped around one job, so the repetitive work runs itself and your people get their hours back. It runs on your own hardware, offline by default. Then we hand over the source code, every key and full control, and step away.
It all runs on your hardware. Cloud AI is added only if you ask for it.
Every screen and workflow shaped around the job your team does today.
Source code, data and keys. Nothing left on our side.
One task, one app — built for a person or a whole company.
You describe the job; we build the application around it and sign it over — on one laptop or across a factory floor.
Individuals
Private tools for your music, health records or creative work that take the chores off your hands, on your own computer or phone.
Corporations
Assistants and workflow engines that take the copying, checking and chasing off your teams, on a private server inside your network.
Factories & technical sites
Instruments connected, inspections and reports automated, every measurement kept on hardware that never leaves the plant.
Anything custom
If it can be specified, it can be built. Bring the problem; leave with a finished system and its keys.
One vault holds everything. AI runs the rest.
Every OddBox system keeps its records, users and activity log as plain Markdown files on hardware you own, and local AI agents do the filing, linking and reporting around the clock.
Your database, in plain files
Records and history as readable notes. Search them, open them in Obsidian, back them up to a USB drive.
Sign-in that stays in the building
Accounts, roles and sessions live inside the vault, with master access held by your safety key. No third-party login.
AI at the controls
Local models create records, answer questions, write reports and flag what looks wrong, so nobody has to do it by hand.
Zero cloud dependency
Unplug the internet and it keeps working. Models, data, users and logs all run on hardware you own.
Five steps, then we’re out of the picture.
No dependency, no backdoor. The last step is us revoking our own access.
Scope
We map the job, the data and where it should run, and agree a clear scope and timeline before a line is written.
Build
We build the software, the automation and the admin console, and wire in the AI models the job needs.
Deploy
We install it on a private server in your building or on your own devices, on locked-down Linux.
Handover
You receive the full source, every credential and a written safety key. We revoke our access in front of you.
Independence
It’s yours. Run it, change it, audit it or have anyone extend it. It keeps working whether OddBox exists tomorrow or not.
Your hardware or the cloud. Your choice.
Your apps and your vault always run on hardware you own. The AI model can run there too, fully offline, or come from a cloud service when you want the largest models without new hardware. Either way, the machine is sized to your workflow: how many people use it, how much they ask and how fast answers must come.
Your own server
A private AI server in your building, or the computer on your desk, with open-weight models such as gpt-oss, Gemma, Llama, Qwen, DeepSeek and Mistral. We size it, install it, tune it and hand it over.
Cloud services
Every major model, including GPT, Claude, Gemini and Grok, which are only offered as cloud services. Ready the same day, with no new hardware, and connected only if you ask for it.
Built for industries where data can’t leave the building.
Fusion labs, ports, refineries, launch sites: places where a data leak is not an option and every hour of an expert counts.
Tell us what you want built.
Describe the job in plain terms: the problem, roughly where it should run, and what “done” looks like. We’ll come back with a clear scope and timeline.
Message sent.
Thanks — we’ll come back with a clear scope and timeline.
Proven platforms. Your choice of AI.
Windows 1126H2
macOS 27Golden Gate
iOS & iPadOS 27iPhone · iPad
Android 17phones · tablets · kiosks
Ubuntu 26.04 LTSResolute Raccoon
Debian 13trixie · 13.7
Linux Mint 22.3Zena
Obsidianlocal Markdown vaults
Ollamalocal model runtime
vLLMGPU inference server
Dockersealed containers
SQLiteembedded database
gpt-ossOpenAI · 20B / 120B
Gemma 4Google DeepMind
Llama 4Meta · Scout / Maverick
Qwen3Alibaba Qwen
DeepSeek-V4Flash / Pro
MistralSmall 3.1 · DevstralOpenAIGPT-6 Astra · Sol · Luna
AnthropicClaude Fable 5.1 · Opus 5.5 · Sonnet 5.5GoogleGemini 3.1 Pro · 3.6 Flash
SpaceX AI (xAI)Grok 4.6Every model above is available as a cloud service. Open-weight models can also run offline on your own server, with hardware sized to your workflow. Cloud AI is connected only if you ask for it — never as a dependency.
Versions as of October 2026. All trademarks belong to their respective owners.