Lower your coding-agent bill by up to 70-80%, same output

No. Claude Code, Codex, Cursor, Cline, Aider and plain SDK calls all keep working. Each developer changes one base URL to the Proxen address and pastes in their own key. Claude Code officially supports routing through an organization's gateway, and the other tools expose the same setting. Prompts, slash commands, extensions and editor integrations stay as they are. The only thing a developer notices is that nothing changed.

Most Claude Code cost is context: the conversation, tool output and files sent with every turn. The proxy trims that to what the model needs, keeps the prompt cache warm so repeated context is billed at the cache rate, and sends easy steps to a cheaper model while hard steps keep the strongest one. Before any rule goes live we test it on your own repos and compare the answers with what your developers get today. You approve every rule, and the log shows which model answered each request.

Every key carries tags, such as a person, a team and a project. You set a monthly budget on any of those: a person, a repo, a client project, an agent or the whole org. Proxen meters every request underneath. At 80 percent of a budget the owner gets an alert. At 100 percent the key stops until you raise the cap or the month resets. Because each budget is separate, one project running hot never drains another. The console shows live spend against every budget so you can move allowance where the work is.

They hit their own cap, not the org's. That is the whole point. With a shared plan one heavy user locks everyone out until the limit resets, and an unattended agent keeps spending until someone notices the invoice. With a key per person and per agent, requests count only against that key's budget. You see who it is in the log, decide whether the work justifies a bigger allowance, and change the cap in the console. Nobody else is affected and nobody has to ask the channel who burnt the budget.

Give each pipeline, review bot or scheduled agent its own key. Work that can wait for results goes through the providers' batch APIs, which Anthropic and OpenAI price at half the normal rate on input and output. We move the right jobs over with you. Interactive keys for developers stay on the normal fast path. On top of that the proxy routes simple tasks to cheaper models and caches repeated context, which brings the cost per request down month after month.

Requests pass through the proxy, which is how it meters and routes them, but by default it does not keep them. Proxen logs who sent the request, which model answered, how many tokens it used and what it cost. Prompt and response bodies are not stored unless you switch that on for your own audit needs. Retention is a setting you choose, and the proxy can run in our cloud or in yours if you want the gateway and its logs inside your own environment. Requests still go out to the AI providers you enable.

Close. Anthropic's own docs describe the supported route for organizations as an LLM gateway: the provider key stays server-side and developers hold gateway credentials instead. Proxen is that gateway, plus the tuning that makes each request cheaper. It runs on API accounts your company owns, so every token is licensed to you, with a budget and a log per developer.