The Proxen proxy
One address for all your AI traffic, tuned to cost as little as it can
Proxen is a proxy: one address that all your AI tools and agents talk to, that we set up and tune for you. We sit with your team, learn your whole setup, build it around how you work and keep tuning it as usage changes, so every request costs as little as it can without hurting the answer. It sends fewer tokens, picks the right model for each task, never pays twice for the same context, and runs work that can wait at half price. For the right workload that means up to 70-80% lower AI cost with the same quality. Your number depends on your workload, and the free bill review tells you what it is before you commit.
the levers
What the proxy does to every request
These levers stack. Each one works on what is left after the one before, which is why a tuned proxy saves far more than any single trick. We pick the mix for your workloads, show you every rule, and you approve it before it goes live.
Fewer tokens
Coding agents resend the whole conversation on every turn, so sessions get more expensive the longer they run. The proxy trims the history and context each request carries to what the model needs, and sets sensible limits on answer length where long answers add nothing.
Lower price per token
Easy requests go to cheaper models. Inside one agent session the proxy picks a cheaper or a stronger model for each task, and keeps the strongest model for the work that needs it. On Anthropic's own price list, Claude Haiku 5.5 costs one fortieth of Claude Opus 5.5 per token on prompts up to 100k tokens.
Never pay twice
Our own cache answers repeated requests without calling a model at all. For repeated context we make full use of the providers' prompt caching: on Claude Opus 5.5 and Sonnet 5.5 a cache read costs 5 percent of the normal input price. We also fix the things that quietly break the cache.
Half price when it can wait
Test generation, review passes, docs and nightly jobs do not need an answer in two seconds. Work like that goes through the providers' batch APIs, which Anthropic and OpenAI price at half the normal rate on both input and output. Interactive work stays on the fast path.
Many models, one address
Claude, GPT, Gemini and open models sit behind one proxy, so a team can use the right model for each job at a lower price without opening another account or learning another tool. You decide which models each team may use.
Built for your setup, not a default config
Every team's bill leaks in different places. A coding-agent team pays for long histories and cache rebuilds. A product team pays a top model to answer easy questions. A data team runs nightly jobs at full price. We read your usage first, then set the levers that matter for you: how much history to trim, which tasks go to which model, what to cache, what can wait for batch. The console shows spend by team, tool and model, before and after.
Same quality as standard, no compromise
A cheaper answer only counts if it is as good. Before a rule goes live we run it against your own work and compare the answers with what you get today. Hard tasks keep the strongest model. You approve every rule, and every request in the log shows which model answered, so anyone can check. If a change makes answers worse, it does not ship.
A key per person and per agent, with a budget and a hard cap
Every person and every agent gets its own key, so savings show up per team and per tool, not as one lump. Each key has a model list, a monthly budget, an alert at 80 percent and a hard stop at 100. Provider keys stay server-side on your company's own accounts. Prompt and response bodies are not stored unless you switch that on.
Any harness plugs in: point it at one address
how we work
Four steps from a high bill to a tuned proxy
Questions about the proxy
Does a cheaper model mean worse answers?
No. The goal is the same quality as standard. Hard tasks keep the strongest model. Easy tasks, like short edits, summaries or commit messages, go to a cheaper model that does them just as well. We test every rule on your own work before it goes live, you approve it, and the log shows which model answered each request.
Where does the up to 70-80% figure come from?
It is what the full set of levers can do on workloads with a lot of room: long agent sessions, a top model used for everything, repeated context and work that could wait. Teams that already use cheaper models or have little repeated context save less. That is why we start with a free bill review on your own usage and tell you your expected number before you commit.
Which tools and harnesses work with it?
Any harness can be integrated. Claude Code, Codex, Cursor, Cline, Aider, the OpenAI and Anthropic SDKs, LangChain, LlamaIndex, n8n and your own agents all point at one address. For most tools that is one base URL change. Claude Code officially supports routing through an organization's LLM gateway.
Do we have to change providers or accounts?
No. The proxy runs on your company's own Anthropic, OpenAI and Google accounts, or we help you open them. You keep the accounts and the history. If you leave, you point your tools back and nothing breaks.
See what the proxy would save on your bill
Send us one month of usage exports from your providers, or just tell us your tools, models and rough monthly spend. We show you where the money goes, what we would change, and the saving we expect, before you commit to anything.