$0
The agent
The app, the local model, and two free skills — Kitchen Timer & Units, and Packing List. No account. No trial that ends.
Local · Offline · No subscription
The gym for your agent
Hydropark is a free local agent — a small AI that runs entirely offline on your own computer, nothing sent to the cloud. Out of the box it can hold a conversation and very little else. Bolt on skills — timers, lists, planners, a switch that turns cups into grams — in whatever combination you want, and the app rebuilds itself around the job. Each skill is $5, yours forever. Save the combination; come back to it on Tuesday. Windows & macOS.
- Base agent
- Free
- Each skill
- $5 once
- Skills at once
- As many as it can carry
- Runs on
- Your CPU
00
You are not
training a model.
Nobody is running gradient descent on your laptop, and anyone who tells you otherwise is selling something. The weights never change. What changes is what the agent is allowed to do — which tools it may call, which voice it speaks in, and what the app looks like while it works.
That is the useful half, and it takes a second instead of a data centre. So think of it less as a dojo and more as an equipment room. You are not building muscle. You are loading the bar.
01
The rack
Every skill has a weight: the prompt it adds, the tools it brings, the panels it draws. Load what you like. The bar will tell you when it has had enough — and it is a real limit, computed the way the app computes it.
Your bench
Plates
Try a routine:
Prompt budget
0 / 3072 tokens
Skills may fill 3,072 tokens; the other 1,024 of the model's 4,096-token window is held back for your conversation, so a long chat never quietly evicts a skill. The budget is the same on every machine — your bench sets speed, not how much fits.
- Verdict
- Ready
- Speed
- 13.0 tok/s
- Skills
- 0
- Tools
- 0
- Panels
- 0
Shared state
Panels the app will draw
The system prompt this builds
Load a plate to see it assemble.
One skill leads and owns the voice. Everything else contributes a single line, its tools, and its panels — which is why a fifth skill costs almost nothing and the first one costs a lot.
Your rack
$0
Both free plates are already in the app. Add a paid one and the price appears here.
02
Save the loadout.
Not the conversation.
A combination that works is worth keeping. Name it and it becomes a template — the same skills, the same lead voice, the same panels in the same places, pinned to versions that behave the way they did the day you saved them.
-
Weeknight Chef
Cooking Assistant leads · Nutrition Coach reads the ingredient list
ingredients -
Trip Week
Travel Planner leads · Packing List writes what you take
packing_list -
Saturday Jobs
Home & DIY leads · Car Care and Garden read the shopping list
shopping_list
Loading a routine is instant and needs no network. It is a small file on your disk that names the skills you already own.
03
House rules
A marketplace of things that reshape your software is only safe if the things cannot do very much. So they cannot.
-
A skill has no code in it
It is a text file: a personality, a list of which built-in tools it may call, and a description of the panels to draw. There is nothing in it that could run.
-
The tools are ours, and there are not many
Start a timer. Convert a unit. Edit a list. Do arithmetic. Do date maths. A skill picks from that fixed, audited catalogue or it does without. It cannot invent a tool, reach the network, or touch a file.
-
One skill leads; the rest assist
Stack four personalities onto a three-billion-parameter model and you get a confused one. So exactly one skill owns the voice — you choose which — and the others contribute a line, their tools, and their panels.
-
The meter blocks, or it warns. It never shrugs
Run out of context and the combination is refused before you enable it, with the name of the plate to drop. Run low on speed and you get an amber warning and the choice. Nothing is silently downgraded behind your back.
-
Skills cannot bring their own stylesheet
They pick from the app's design tokens. Twelve skills by twelve authors still look like one program, and every panel is keyboard-operable whether its author thought about that or not.
-
Nothing goes on the floor untested
Before a skill is published it has to pass a behavioural exam on the reference model — representative prompts in, expected tool calls and panels out — plus a safety review and a hard budget on how much prompt it is allowed to spend.
04
The equipment
lives in your house.
The model is a two-gigabyte file on your hard drive. There is no session to renew, no server holding your conversation, no monthly bill that decides whether today is a day you get to cook. Turn the Wi-Fi off and everything above still works — chatting, every skill you own, every routine you saved.
What you bought is checked against a signature, not a clock. There is no expiry date and no phone-home, so a laptop that never sees the internet again keeps its plates. And we have written into the terms what happens to all of it if this company ever closes: a final offline bundle, your skills, permanent licences, and never a released signing key.
- Conversations never leave the machine
- No account needed to use it
- Qwen2.5-3B, Apache-2.0, via llama.cpp
- Windows and macOS
05
There is no
membership fee.
The gym is free. The bar is free. Two plates are free. You buy the plates you want, once, and they are yours the way a book is yours.
$5
A plate
Any single skill, one payment, perpetual. Install it on up to five of your machines. Uninstall and reinstall for free, forever.
Most people start here$18
A rack
Five skills in one go, versus $25 apart. A themed three-pack is $12. Or top up a wallet once and spend it without a checkout each time.
Prices shown in US dollars; your local price and tax appear at checkout. Fourteen-day refund, no questions. No subscription, no seats, no usage limits — there is nothing to meter, because it is your processor doing the work.
06
Questions a
spotter would ask
So I really can't train it?
Not in the sense the word usually means. The model's weights are fixed and identical on every machine. A skill changes the instructions it runs under, the tools it can reach, and the interface you see — not what it learned. We call it a gym because you are loading equipment onto something, not because there is a training loop. If you want a model fine-tuned on your data, this is the wrong product and we would rather say so here than in a refund email.
How many skills can it hold at once?
There is no fixed number, which is why there is a meter. Each skill spends part of the model's context window — its instructions, its tools, and its panels — and the app holds a fixed reserve back for your live conversation. When the next skill would not fit, the planner refuses it before you enable it and names the plate to drop. That budget is the same on every machine: a stronger bench buys faster replies and fewer speed warnings, not more plates. Toggle the bench above and the speed readout moves while the capacity stays put — that is the real rule, not a marketing number.
Why does one skill have to lead?
Because concatenating four personalities onto a three-billion-parameter model produces something with none of them. The lead skill's full instructions go in; every other skill contributes one line describing what it can do, plus its tools and its panels. You pick the leader. The order is deterministic, so the same loadout assembles the same way on every machine, every launch.
What happens when two skills disagree?
Some pairs are simply refused — the free Kitchen Timer and the full Cooking Assistant are the same muscle at different sizes, so the app will not let you carry both. Where two skills touch the same list, exactly one of them owns it and the other reads or appends; the app decides, deterministically, and tells you which. Panels that are genuinely identical get merged rather than drawn twice.
Can I write my own skill?
Not yet, honestly. The catalogue is first-party and curated so that the safety story above is actually true. A closed-beta SDK goes to a handful of design partners first, and a public one — with the same certification pipeline and a revenue share — comes after that. We would rather ship a small shelf that works than a large one that does not.
Is it as good as ChatGPT?
At answering anything, no, and it is not pretending to. It is a small model on your processor. What it has instead is a timer that is already running, a list that stays on screen, a unit switch that works mid-sentence, and no internet connection. For a structured job you repeat every week, that is a better tool. For an essay about the Peloponnesian War, open a browser.
The bar is empty and it is free.
Windows & macOS · ~2 GB model, downloaded once · no account