ModelPaddockFor teams & IT
Your developers already use five AIs. Give them one safe front door.
ModelPaddock puts Claude Code, Codex, Gemini, goose and on-device models behind one window, and keeps secrets, spend and risky commands in check. It runs on each laptop and has no cloud of its own.
No account, no telemetry, no new vendor in your data path.
Built to pass a security review
Every feature below is in the app today. Nothing on this page is a promise about the future.
Data stays on the laptop
Search, costs, ratings and history are built from files already on the machine and never uploaded. ModelPaddock only starts AI tools your people already use, with the accounts they already have.
Leaks stopped before they leave
Privacy autopilot spots API keys, passwords, card numbers and bank numbers in a prompt. By default that prompt goes to an on-device model instead of a cloud AI. It can also mask the private parts or ask first.
A person approves risky steps
When Claude Code asks to run a command, the developer can allow or deny it from their phone. If nobody answers, nothing is allowed; it never says yes by itself. Overnight Night Shift work lands on its own git branch, so nothing is merged or pushed without review, and each job is checked first: the project’s tests (sealed off from the network where the system supports it), plus a read-only review by a second, different AI.
Every AI edit can be undone
Rewind saves a snapshot before an AI started from ModelPaddock edits a project, in a private history that never touches the team’s own git repository. Collision radar warns when two AIs work in the same project at once.
AI spend you can see
Costs & limits shows tokens and estimated spend per account and model, with budgets and a run-out forecast. “What should I use?” picks the cheapest option that is good enough, including free on-device models.
A record of what each AI did
The AI team room keeps one log per project: which AI did what, what it handed off, and what it asked. The next AI (or the next person) starts with context instead of guesses.
Hardware stays safe
On-device models are paused while a laptop is hot and stopped before it overheats, so local AI doesn’t cost you a machine.
Everything ModelPaddock connects to
Most vendors make your security team ask. This is the complete list, from the code, so your reviewer can check it with a network monitor.
| Connects to | Only when | What is sent |
|---|---|---|
Ollama on the same computer127.0.0.1:11434 | Someone uses an on-device model | The prompt. It never leaves the machine. |
Your phone on your own networkport 8765 | “Pair your phone” is switched on (off by default) | Status, approve/deny answers and (if allowed) Night Shift jobs, protected by a PIN. Over Wi-Fi or the phone’s Bluetooth network. Requests for other host names are refused. |
| A notification address you choose for example ntfy.sh | Someone enters a push URL (empty by default; https only) | A short title and line of text, such as an approval request or the Night Shift report. |
OpenRouteropenrouter.ai | Someone adds their own OpenRouter key | Race prompts sent to cloud models they chose, and a balance check. No key, no connection. |
Lemon Squeezy (license store)api.lemonsqueezy.com | A store-bought license is activated or checked | The license key and “ModelPaddock on” plus the system type (Linux, Mac or Windows), never the computer’s name. Offline license keys need no connection at all. |
| An AI tool’s official download e.g. ollama.com, github.com | Someone clicks Install for a missing tool, after seeing the exact command | A normal download of that tool. ModelPaddock adds nothing to it. |
| Your AI tools Claude Code, Codex, Gemini CLI, goose… | Someone opens one | Nothing from ModelPaddock. Each tool connects to its own provider exactly as it would without ModelPaddock. |
| Nothing else. No analytics, no telemetry, no crash reports, no automatic update checks, no account. | ||
to attach it to your security review.
Estimate it for your team
Put in your own numbers. This is simple arithmetic on what you enter, not a promise.
Assumes 21 working days a month. On-device models cost nothing per prompt but are slower; Privacy autopilot already sends prompts with secrets on-device.
For the people who roll it out
Works today
- One settings file. Everything lives in a plain TOML file (
~/.config/aihub/config.tomlon Linux), so you can pre-fill privacy mode, budgets and providers before handing out laptops. - No admin rights on Linux. The AppImage runs from the user’s own folder.
- Clean removal.
modelpaddock --uninstallremoves everything it added;--purgealso deletes settings and data. The developers’ AI tools and chats are never touched. - Offline licenses. Signed license keys are checked on the machine, so they work behind firewalls and on air-gapped networks.
- Uses accounts you already govern. ModelPaddock adds no AI provider of its own. Your existing Anthropic, OpenAI or Google agreements still apply.
Not there yet honest list
- No central admin console or single sign-on. Each laptop is set up on its own.
- Licenses are per person; there is no seat management yet.
- Mac and Windows builds aren’t code-signed yet, so your device management may block them.
- Linux has the most features today (best on KDE Plasma).
Tell us which of these blocks you. Company requests decide what we build next.
Talk to us about your team
Tell us a little about your team and we’ll reply about a pilot and pricing for 5 or more people.
We keep your work email, company, team size, the system you picked, your note, your country and the time, only to reply to you. Ask us and we delete it.