Example setup: a 60-person engineering team
Illustrative example, not a customer. How a 60-person engineering organization could put Claude Code, Codex and Cursor behind one gateway with per-person keys, model lists per squad, and batch pricing for nightly jobs.
Overview
Illustrative example, not a customer. Numbers are the setup values, not results.
A scale-up with 60 engineers in eight squads, plus a platform team that runs CI and internal tools
A mix of personal Max seats on expenses, a few raw provider keys in CI secrets, and no view of spend by squad
Company API accounts, 60 person keys and 9 service keys, one team per squad, routing and caching rules, batch for nightly jobs
Challenge
In this example, engineers are already using Claude Code, Codex and Cursor every day, paid for through a mix of expensed personal seats and a couple of raw provider keys that ended up in CI secrets. The engineering lead cannot see spend by squad, security cannot see who holds a provider key, and a weekly usage limit on one shared seat has locked people out twice.
The platform team also runs nightly test generation and evaluation jobs that do not need to be interactive but are billed at the full rate.
setup values
Key per engineer and per CI service. Raw provider keys removed from every secret store
Squad teams, each with its own model list and monthly budget
Batch discount applied to nightly jobs that can wait up to 24 hours (OpenAI and Anthropic batch docs)
Challenge
Illustrative example, not a customer. Sixty engineers, three tools, personal seats on expenses and provider keys in CI. No spend by squad, no record of who holds a key, and lockouts when a shared seat hits its weekly limit.
The company's own Anthropic, OpenAI and Google API accounts go into Proxen. Eight squad teams are created, plus a platform team. Each engineer gets a key in their own name; each CI pipeline and internal tool gets a service key with a narrow model list. The raw provider keys are rotated at the provider and removed from every secret store, which is the change security cared most about.
Engineers change one base URL and one key in Claude Code, Codex and Cursor. Claude Code officially supports routing through an organization's gateway, so this is the documented configuration. The first request from each engineer is checked in the spend view under their own name.
Setup
One gateway address for every tool. One key per engineer and per service, grouped by squad. Trimmed history, approved routing rules that send simple work to cheaper models, caching for the repeated project context, and nightly jobs at the batch rate.
Each squad gets a model list and a monthly budget. Each squad approves routing rules that send commit messages, reformatting and summaries to a cheaper model, while the strongest models stay on design and debugging work. Long agent sessions get their history trimmed, and quality is checked on the squad's own work before each rule goes live. Caching rules cover the repeated project instructions and file context that every coding-tool request carries.
The platform team moves nightly test generation and evaluation runs, which can wait for their results, to the providers' batch endpoints, which both OpenAI and Anthropic price at a 50 percent discount for jobs that can wait up to 24 hours. Alerts at 80 percent go to the engineer and the squad lead; the hard stop at 100 percent means a long agent run cannot take the whole squad's allowance. Personal seats come off expenses once every engineer's requests are arriving under their own key.
built on evidence
Every setup example rests on the same four facts about how Proxen works.
Key per person, with its own budget and log
Shared passwords left once rollout is done
Base URL to change in Claude Code, Codex, Cursor or your SDK
Batch discount OpenAI and Anthropic give on work that can wait
Want a setup like this for your team?
Send one month of usage and we will map a setup like this to your accounts, with the saving to expect.
