BYOK vs an account: which one should you pick for AI coding?
When to bring your own API keys, when an account pays for itself, and why Ava Supernova supports both on every plan — including Free.
Every AI coding tool eventually asks you the same question: do you want to use our models, or bring your own? The answer shapes what you spend, how private your code stays, and which models you can actually reach. Ava Supernova supports both on every plan, including Free — this guide is the decision tree we wish we had when we were picking.
What "BYOK" actually means
Bring-your-own-key means you hold the API key from a model provider — DeepSeek, Qwen, Kimi, Mistral, Zhipu, MiniMax, Xiaomi, Tencent Hunyuan, NVIDIA, or anything OpenAI-compatible including models running on your own machine — and Ava uses your key to make the call. You pay the provider directly. We charge nothing for the request itself.
An account means we hold the keys, run the models on our own infrastructure, and bill you in credits. You trade per-request fiddling for one predictable price and a lineup that is already tuned and tested.
What an account actually gets you: two fleets
This is the part most comparisons miss. An account is not one model behind a logo — it is two complete routing fleets, and you pick which one runs your work. Both are live on every plan, including Free.
- Maestro — one conductor handles every persona and every step, Qwen end to end. Production-tuned and the default, for daily work, predictable cost, and anyone who wants Ava to just work without thinking about routing.
- Supernova — a polyglot ensemble. A DeepSeek Flash coordinator picks the best specialist for each task, with Qwen 3.7 Plus taking Builder and vision. Built for heavy multi-step work where each subtask wants its own specialist.
Assembling either of those yourself means holding multiple provider keys, deciding which model handles which step, and maintaining that as models change underneath you. The fleets are the work we have already done so you do not have to.
Pick BYOK when
- You already have paid credit sitting at DeepSeek, Moonshot, Mistral or another provider that would otherwise go unused.
- You want a specific model we do not serve on an account — Kimi is the usual one. New releases often land on BYOK first.
- Your code is sensitive enough that you want every request going straight to the provider, with nothing of ours in the middle.
- You are running models locally and want nothing leaving the machine at all.
- You are heavy enough that provider list price beats any plan.
Pick an account when
- You want one predictable bill instead of six, and a cap on surprise costs.
- You want the routing handled — a cheap classifier for the easy calls, a strong model for the heavy lift — without managing it yourself.
- You want Creative Studio: image, video, voice and design generation, which is genuinely painful to assemble provider by provider.
- You would rather spend the evening building than tuning a model lineup.
You do not have to choose
Both run side by side on any plan. Your account credits handle everyday agentic work through whichever fleet you picked, and Creative Studio alongside it; your own keys take over whenever you want a model we do not serve, or when a piece of work should never touch our infrastructure. BYOK requests never touch your credit allowance — they go straight to the provider.
What about privacy?
BYOK is the stricter path: the request goes straight to the provider and we never see the payload. Local models are stricter still — nothing leaves your machine. With an account we do handle the request, because we have to in order to route and bill it.
What does not change is where everything else lives. Ava is local-first by design: your memory, your chat history and your generated files are written to your own machine first, on every plan and either path. That is the floor, not a setting. There is no cloud sync: none of it is uploaded, on any plan. The only thing that travels on the managed path is the conversation a model needs in order to reply. And your code is never used to train models.
The short answer: start on Free, use both, and let your own usage tell you which way to lean. Your keys, your data, your choice.