Bring your own API key: what building with Adrian actually costs
· updated
There are three ways to pay for building with Adrian: nothing at all on local models, your provider's rates on your own API key, or a monthly subscription for hosted agents. Two of the three are free, and which of those two suits you comes down to your hardware. This is the version without the pricing-page pressure.
The three ways to pay
- Local models through Ollama — free. No account, no key, no bill. How that works in practice.
- Your own API key — you pay your provider directly, at their rates. Adrian takes no cut and adds no markup.
- Hosted Adrian Agents — a monthly subscription (Pro, Max or Ultra) with an allowance of Adrian Credits included.
The first two are the free desktop app. There is no trial timer on them and no feature that unlocks only if you subscribe to something.
BYOK: you pay the provider, and only the provider
Bring your own key and Adrian is a client for a model you already pay for. Paste a key from OpenAI, Anthropic or another supported provider and your usage is billed by them, on their pricing, to your account. Nothing is proxied through us and no percentage is skimmed on the way past. The key stays on your device rather than on our servers, which also means we cannot see or meter what you spend.
The honest catch is that a provider bill is variable by nature. A long debugging session on an expensive model costs what it costs, and the meter you are watching is theirs, not ours. If a predictable monthly number matters more to you than a low one, that is the argument for a plan rather than a key.
What actually drives the bill
A build system spends more tokens than a chat window does for the same question, and it is worth understanding why before you are surprised by an invoice. Four things move the number.
- Context. Generating a change to an existing project means the model reads the relevant files first. A bigger project is a bigger prompt, on every pass.
- Review. Adrian re-reads the code it just wrote and checks it against the request, so a build pays for generation and then again for inspection.
- Repair. When the compile gate or a check finds something, the fix is another round trip. A build that needed two repairs cost roughly three builds' worth of generation.
- Iteration. This is the big one, and it is you rather than the system. Twenty follow-up prompts is twenty builds, and follow-ups carry the project's context with them.
None of that is unique to Adrian — it is what agentic building costs anywhere. What is worth knowing is that the multiplier is on your side of the meter under BYOK, so the lever that matters most is model choice, not tool choice.
You are paying for two roles
Adrian assigns a coder and a designer separately, and the assignment is visible and changeable. That is a cost lever most people miss: the design pass does not need your most expensive model, and pointing the two roles at different models — or running the designer locally and the coder on a key — is a legitimate way to cut the bill without giving up the part you actually care about.
Local: free, paid for in hardware and patience
Running through Ollama costs nothing per build, forever. What it costs instead is machine capability. On a workstation with a decent GPU this is the best deal in the list. On a thin laptop it is slow enough that you will feel it on every iteration — free, but not costless.
The trade is real in both directions. Local is the only option here with a fixed cost of exactly zero no matter how many times you go round the loop, which is precisely the scenario that makes a metered bill uncomfortable. If you iterate a lot, slow and free can beat fast and metered. Which model fits your machine covers what your RAM actually supports.
Hosted Adrian Agents
Paid plans meter hosted agent usage in Adrian Credits, shown as an allowance bar in the app. Every plan has a monthly allowance sized so you always know where you stand, and you can cancel at any time. The pricing page has the current numbers, and subscriptions are processed by Paddle.com, our Merchant of Record.
What a plan buys that a key does not is a ceiling. The allowance bar is a number you can see going down, which is a different experience from a provider dashboard you check afterwards. What it costs you is optionality: a key lets you switch models the day a better one ships, and a plan runs the crew we assembled.
Which one to start on
Start on the free desktop app. If you already hold a provider key, use it — you will get frontier-model quality at exactly the price you already pay. If you do not, and your machine can hold a reasonable model, start local and find out how far the build system carries a small model before you spend anything at all. Move to a plan when you have decided Adrian is worth a monthly line item, not before.
The order matters more than the choice. All three run the same pipeline, so anything you learn on the free path transfers, and nothing you build locally has to be rebuilt if you later switch. Starting at the paid tier is the one route that asks you to decide before you have any evidence.