One key. Every app. Yours.

The current model is broken

Every agentic app ships its own AI integration. ChatGPT-for-X, Claude-for-Y, Cursor, Replit, the seventeen new note-taking apps that all do the same thing. Each one charges twenty to a hundred dollars a month. Each one sits between you and the model, logs your conversations, owns your usage history, and prices its access at two to three times the underlying API cost.

If you use five agentic tools, you're paying a hundred to five hundred dollars a month. Most of that is markup. The actual compute is cheap. What you're paying for is the privilege of accessing the model through someone else's plumbing.

You can't use your conversations across apps. You can't take them with you. When an app pivots or shuts down, and most do, you lose everything. You have no leverage to switch, because switching means losing your history. The apps know this and price accordingly.

The inverse

The model that makes sense is the one we already use for the rest of the internet. You don't pay every website to handle your login. You have a password manager. You don't pay every email service for SMTP. You have an account that any client can use. The infrastructure layer is something you own; apps wrap around it.

Bring Your Own Key. You hold one API key: Anthropic, OpenAI, whoever you trust. Apps are clients that authenticate to your key. The compute is yours. The data is yours. The cost is whatever the model provider charges, no markup.

Apps compete on UX and intelligence layered on top, not on extracting rent from you.

Why it matters

Economics. Cost-per-prompt at the model layer is small. When five apps each charge twenty dollars and each layers markup, you're paying ten times what you should. BYOK collapses the stack: you pay the model provider directly, often five to twenty dollars a month for moderate usage, and apps charge a small license fee for their UX. Total cost: a fraction of today.

Privacy. With managed-model apps, your data passes through the app's servers, gets logged, may be used for training, sits as conversation history on someone else's box. With BYOK, your data goes app → your key → model. The app doesn't need to keep your conversations. It often doesn't.

Lock-in. When an app shuts down, and most do eventually, you lose the conversation history, the custom prompts, the integrations. With BYOK, your conversations can live locally. Your prompts are portable. Your integrations are config you own. The app is a tool, not a vault.

Pace of innovation. Today's model rewards apps that own the entire stack: model partnership, GPU contracts, billing infrastructure, auth, support. That favors capital-rich incumbents. BYOK lowers the bar dramatically. Anyone with a good UX idea can build it. The interesting products of the next decade will come from people who don't need to own the model.

User leverage. With BYOK, you can switch freely. Use multiple apps for the same task. A/B test. Leave. Apps have to compete on actual UX value, not on holding your data hostage.

What's missing today

The plumbing for this isn't there yet. There's no OAuth-equivalent for AI keys. No clean way for an app to request scoped, time-bound, budget-capped access to your model provider. No standard wallet UX for managing keys across apps. Audit logs are non-existent.

These are solvable problems. They're not particularly hard. They're just not what gets built when the entire industry's incentives point toward managed-model lock-in.

What we're exploring

Madmaxlabs is building toward the inverse. Every agentic app we ship will work this way:

• BYOK by default. No managed model. You connect your key. We never see your conversations.
• Local-first storage. Your conversation history lives on your device. Export anytime, in formats that aren't ours.
• Budget caps per app. The app declares its expected spend. Anything beyond that is a hard stop.
• Portable prompts. Custom system prompts and workflows you've tuned move with you.

Over time, the wallet itself. The thing that holds your keys, brokers app access, shows you costs in real time, lets you revoke anything with one tap.

We're early. The first apps are notes-and-thoughts territory. The vision is bigger.

The bigger bet

BYOK is one instance of a broader move toward user-owned infrastructure. The pattern is older than agentic software: email works this way, browsers work this way, password managers work this way. The user owns the substrate; apps are clients. The thing that's been broken in the last decade of consumer software isn't whether users could own the substrate. It's that platforms got there first.

Agentic is the first place that's shifting back. A few other instances we're betting on:

• Local-first data. Your conversations, your ledger, your photos, your notes. Held on your device, with apps as views into them. Cloud sync is a feature you opt into, not the architecture.
• On-device intelligence. Vision and language models running locally for the things that don't need a cloud. Privacy improves. Latency improves. The economics shift.
• User-owned identity. The wallet metaphor works for keys. It also works for who you are online. Not just an API key but a passport.
• Portable workflows. The prompts you've tuned, the configurations you've built up. They travel with you, stored where you decide.

What ties these together: the user owns the substrate. Apps compete on what they do with it.

This is a bet, not a prediction. The alternative is managed AI as the next platform layer. That's also possible. We think the user-owned version wins because the economics are better for the user, the tooling is becoming feasible, and the trust required for the platform version is fragile.