Everything public about Nexara.

Your conversations and files belong to you. We don't read them, we don't sell them, and we don't train on them. Chats are saved to your account in Supabase so you can pick them up on any device, and memory features only ever keep what you explicitly choose.

The people running Nexara don't browse conversations. What reaches us during normal use is technical usage only — typically a webhook notice like "user just used GPT-5.6 Sol" so we can watch costs and reliability. No message text or file content rides along with those notices, and no other data about your activity is saved.

Your name and email only ever leave your account if you send us feedback or report an error — and even then they are never shown publicly. They stay with the small, trusted team that fixes the issue, only for as long as the fix needs them.

The short rule is the same as with ChatGPT, Claude, or Gemini: if it's a secret, don't paste it. Passwords and API keys, private documents, personal information (ID numbers, addresses), proprietary code, and confidential business information should stay out of any AI chat — including Nexara's. To answer you, the relevant parts of a conversation travel to the model provider for that request (the Privacy Policy spells out who sees what); a secret pasted into a chat is a secret that has left your control.

Need help with sensitive code or documents? Paste a redacted version with the real secrets replaced by placeholders — that's enough for almost every question.

Compute is Nexara's currency: 1M Compute = $0.80, and every model bills from that pool at its own provider rate. Settings → Compute shows usage percentages and reset timing.

To make it concrete (rough figures — they vary with model and message length): a short casual reply usually costs a few hundred to a couple of thousand Compute; a long coding turn that builds something big can run tens of thousands; one Image Studio picture is exactly 12,500. So a few million Compute is hundreds of everyday messages — or roughly ten images plus casual chatting — and it goes fast under heavy agent/IDE sessions.

Free users receive a daily Compute allowance to explore Nexara; guests receive a smaller allowance. Paid plans provide weekly Compute worth 1.5× plan price (2× on web chat).

Full models — the real thing. Every plan can use the entire catalog, and what you get is the provider's actual model with its full context window and reasoning modes. Nexara never serves secretly smaller, quantized, or watered-down stand-ins; names like "Lite" or "Flash" in the catalog are the providers' own model names, not downgrades we apply.

Plans don't restrict which models you can pick — they only change how much Compute you have to spend. The free tier gets the same full models with a small Daily Compute limit. If a provider model is briefly down, a request may fall back to another capable model, and the chat always tells you when that happens.

Mostly — but we'll be straight about it: Nexara is run by a small team, not a company with giant infrastructure budgets. When a lot of people are online at once, replies can slow down, and there are real limits (the Daily Compute limit and Weekly Compute limit) that keep the service from being swamped. We scale what we can within our budget and watch load continuously.

In normal use it stays usable. If we ever hit a rough patch, you'll see it acknowledged honestly in the Updates log rather than hidden.

They're real and people use them every day — file-building artifacts with live previews, the agent loop, websites and slides, NexaraCode, the CLI, and the desktop and Android assistants. We wouldn't claim we're the best at everything; bigger tools with bigger teams exist. What we can promise is that we genuinely try to make ours good and keep improving them.

The IDE and NexaraCode are still in Beta — you will hit rough edges. Try everything yourself, and if something is weak, tell us exactly what was weak: feedback is the roadmap.

That's the plan. We intend to keep Nexara running for a very long time. Realistically, staying up long-term depends on Nexara getting enough support (plans, donations) and feedback to cover the model bills and keep the team shipping.

As long as that support keeps coming, we plan to keep going — and if things ever change, you'll hear it from us clearly and in advance, never find out after the fact.

Automatic renewal? Yes — paid plans are subscriptions billed automatically on the interval you chose at checkout. Monthly renews monthly; yearly (−20%) is charged upfront as one payment and renews at the end of the year.

Cancelling is easy. Turn off future renewals any time from your billing controls — no contracts and no cancellation fees. Cancellation takes effect immediately: the paid plan is removed and your account drops to the free tier. Your chats, files, Library and personas stay exactly where they are. Any unused paid-period time is not described here as a refund; contact support if you need help with a billing question.

Unused Compute & refunds. You're never charged past the period you paid for. Refunds for unused plan time follow the terms shown at checkout and applicable law — if you have a question, send it via feedback or report-an-error and a real person on the small team will sort it out with you.

Payment processor. Checkout and subscriptions run through Tebex, a reputable, widely used processor — card details go to Tebex, and Nexara never sees or stores your full card number.

It depends on the surface, and the honest picture is: the web chat is sandboxed; the on-device tools are powerful by design.

In the web chat an agent sees your conversation and anything you attach, and it can create artifacts — files, HTML pages, images, games — that render inside the chat and that you download. It cannot reach your device's filesystem or run commands on your computer from the website. Web lookups happen only through an explicit, visible web-search step mid-conversation, and memory features save only what you choose to keep.

Connected services. Automations, browser actions, and integrations only work with accounts you explicitly connect, and they act under the permission you grant when connecting. Don't connect anything you wouldn't want an agent touching.

On your device the desktop and Android assistants and the IDE/CLI harness deliberately have more power: they can read your screen, open apps, toggle device settings, and in NexaraCode and CLI sessions read, write, and run commands inside the working folder you open. That's the point of an on-device agent — and those runs happen only in sessions you start yourself on your own machine.

The universal rule: never hand any agent a password, API key, or secret. Keep secrets in places only you control, and check what a session can reach before letting it loose on important files.

Short answer: yes, you could move tomorrow without losing your work. Nexara doesn't hold what you create hostage.

Everything the AI makes is ordinary, open stuff you already own: website and code artifacts download as files or a .zip, documents and slides download in their standard formats, images save to your device or Library, and 3D models export to .glb/.obj/.stl. Nothing is encrypted to your Nexara account or stored in a proprietary format that only Nexara can open.

Chats are plain text on your account: you can copy any message or artifact out at any time, and share links exist for anything you want to publish. There isn't a one-click "export every chat" button yet — treat Nexara like any tool and download what you care about. Account deletion is handled through in-app support under the Privacy Policy, and if you ever cancel you keep your account, chats, and anything you already downloaded.

They're identifiers in our model catalog, not claims that OpenAI or Anthropic publish those exact retail names. The catalog follows the standard router convention — provider/name — the same format aggregators like OpenRouter use, and the prefix simply names the provider family a model belongs to.

Every model is served through the real provider pipeline: either a direct provider connection (Claude models through Anthropic, GPT models through OpenAI) or an open model router, depending on the model. No model is a custom stand-in we trained — you get the full, unmodified model behind each catalog entry, and the chat tells you exactly which model served your reply.