A lower AI bill your security team can sign off

Not through us. Proxen does not use your prompts, responses or logs to train any model. Requests go to each provider under your company's own API accounts, so the provider's business API terms apply, not consumer terms.

Not by default. For every request we log four things: who sent it (the named key, whether a person or an agent), which model answered, how many tokens went in and out, and what it cost. That is enough to attribute spend, spot misuse and answer an audit. Prompt and response bodies pass through and are not kept unless your admins switch body logging on for a specific team or key. Retention is a setting you choose, and the log can be filtered in the console for any review.

Your Anthropic, OpenAI and Google keys belong to your company. They are stored on the Proxen server, encrypted at rest, and used only to forward requests. No team member sees them after setup, and they never appear in a tool, a repo, an agent config or a CI variable. People and agents hold Proxen keys instead. If you run the proxy in your own cloud, the provider keys never leave your account at all.

In our cloud or in yours. The proxy can run in our cloud, in your own cloud account, or on servers you manage. In your own environment the gateway and its data stores, including the provider keys and the logs, run there, which is usually what client contracts and internal policy require. Requests still go out to the AI providers you enable. In your cloud, we deploy it with your platform team in the way you already run software, and model lists, budgets and routing rules are managed the same way either side. Hosting is agreed at setup and can change later.

Your admins see the request log for your own keys. Our staff access it only to run the service or to help with a support request you raise. When an engagement ends, we delete your keys and logs on request.

Immediately. Open the person or agent in the console and revoke their key. It stops working on the next request. Because provider keys were never on their machine and nobody else shares their key, only that one key needs replacing and nobody else is interrupted. The log records the revocation time. For contractors you can set an expiry date at issue, so access ends on schedule without anyone remembering to act. If you already run an offboarding checklist, revoking the Proxen key becomes one line on it.

Every key has a monthly budget, an alert at 80 percent and a hard stop at 100. The check happens before the request goes to the provider: once a key stops at its budget, the next request is refused before it reaches the provider. Spend is set up front, not discovered on the invoice.

A raw provider key has no budget, no model list and no per-person log on your side. Anyone holding one can spend without limit until the provider's own billing alerts fire, which is how one leaked Gemini key reached $82,314 in 48 hours (The Register, March 2026). A Proxen key is scoped: it can only reach the models you allow, it stops at a hard cap, it is revoked on its own, and every request it makes is logged against a name. Anthropic's own gateway documentation describes this pattern: the provider key stays server-side and people hold gateway credentials instead.

In practice four things: who may use AI (named people and named agents, not shared logins), which models each team may reach, what may be sent and what is kept, and who owns the spend. Proxen turns each of those into a setting: a key per person and per agent, a model allow-list per team, metadata-only logging with retention you choose, and a budget with a hard cap on every key. The policy stays a document; the proxy is what enforces it on every request.

Yes, because it is one of the ways the proxy cuts cost. There are two kinds. Provider prompt caching happens at Anthropic, OpenAI or Google under your own account. Our own cache answers repeated requests at the proxy, so it holds request and answer data for a short time. Which teams use it, how long entries live and where it runs (our cloud or yours) are agreed with you and written into your security answers.

Proxen is run as critical infrastructure, in our cloud or next to your other critical services in yours. Each tool connects through one base URL that your admins control, so access is never locked to us: your platform team decides how traffic flows, and we agree how the gateway is run and monitored with you at setup.

Proxen speaks the API formats your tools already use, including the Anthropic, OpenAI and Gemini APIs, so Claude Code, Codex, Cursor and your own SDK calls connect by changing one base URL. Responses stream back token by token, the same as from the provider.

No. Their tools work the way they did before, with the same commands, the same models and answers streamed back token by token. The budget check happens in line, so the speed your developers feel is the AI provider's own.

Tell us what your company requires. Enterprise requirements such as single sign-on, user provisioning, audit log export and where data is stored are scoped into your engagement, and we confirm each one in writing before you sign.

Contact and newsletter details are sent to the systems we use to receive and manage enquiries. Ask us for the current processor list and the deletion process.

Yes. Send it to us and we answer it in writing, question by question, before you sign anything, including how each of your requirements is met.