Start with one base URL: everything TrustedRouter does
If you have never heard of TrustedRouter, here is the whole thing in one sentence: it is an OpenAI-compatible API that routes your request to hundreds of models across dozens of providers, it keeps no prompt or output logs, and it can prove that claim with a hardware attestation you check yourself. Everything below is a map of where to go next, because we have written a lot and a new person has no idea which door to open first.
Start with the agent base URL. That is the entire migration. Point your client at https://api.trustedrouter.com/v1, keep your SDK, keep your code, and you are done. Claude Code, Codex, Cursor, eval runners, anything speaking the OpenAI SDK — one environment variable. If you are coming off OpenRouter there is a migration page that is shorter than this paragraph. I wrote about how little work this is in Migrating to TrustedRouter is one URL, and the OpenAI-compatible API page lists exactly which endpoints exist.
The privacy claim is checkable, which is the unusual part
Every router says it does not log your prompts. That is a promise, and a promise is worth what the company is worth. We run the gateway inside a secure enclave, TLS terminates inside that enclave, and the enclave signs a statement about which exact code image is running. You fetch it. You compare the digest to the open source build. If they match, the software handling your prompt is the software you read. The argument is in Attestation is all you need and the shorter version is on the no-log is a promise, attestation is proof post. Go check it live at trustedrouter.com/security. If you want the mechanics of enclaves and why they matter for inference, how confidential computing protects AI prompts is the explainer.
The gateway is one hop. The provider on the other end is a second one, and it has its own retention policy. So we published it. The providers page compares every upstream on retention, region, confidential compute, encrypted routes, and measured latency, and it is machine readable at https://api.trustedrouter.com/v1/providers. If you need zero data retention, filter for ZDR and route only to those. If you need more, route to a provider running its own enclave and you get privacy all the way down to encrypted GPU memory. Some providers are still training on what you send them, which is the subject of they are still training on your data. Read that one before you pick a default.
Europe gets its own gateway
If your data cannot leave the EU, use the EU gateway. There is a Europe West region at https://api-europe-west4.quillrouter.com/v1, an trustedrouter/eu alias that prefers EU and privacy-forward routes, and the same attestation you can verify on the main gateway. The legal paperwork people actually ask for is up too: a DPA, a BAA, and a subprocessor list. Nobody enjoys reading those pages. Everybody's procurement team asks for them.
Then the boring, useful question: which model?
Smart, cheap, fast — pick two. Model comparisons puts two models side by side on context, price, privacy posture, and the routes that actually serve them, using current catalog data rather than a screenshot from March. Choose takes a description of your task and how private it has to be and picks for you, and your coding agent can call it through MCP. The full catalog lives at models, prices at pricing, and measured route performance at the leaderboard and provider latency benchmarks. Two posts worth your time here: how to choose a model, and the best open models aren't on your leaderboard.
Model choice has a censorship dimension most benchmarks skip. We measured which models refuse and where the refusal comes from, and often it is the host, not the model — censored at the host, not the model, plus the models that go silent on China and the models that say no. If you are routing customer traffic, knowing which routes go quiet on which topics is operational information.
The rest of the surface
Once the base URL works, the features you will want next are prompt caching, batch, web search, evals, request tagging for per-team cost attribution, video models, and x402 if you want an agent paying per request. There is Synth for fusing several models into one answer, and ten cheap runs beat the frontier is the result that made me take it seriously. Sign in with TrustedRouter lets your users bring their own keys. Uptime is public at status, and how we hold it is in achieving 99.999% uptime as a startup. For agents that read docs instead of pages, everything is dumped at llms-full.txt.
The reasonable objection is that a router is a middleman, and adding a middleman to your most sensitive traffic is a bad trade. It is, when the middleman is opaque. Ours is open source end to end — gateway, config, bring-up, UI — so you can read what runs, verify what is running, and if you would rather not trust us at all, run it yourself on your own hardware. That option is the point. I want a hundred routers like this to exist, and I wrote about why in one API, all the LLMs, provably private and why not OpenRouter, LiteLLM, or Vercel AI Gateway.
Thirty seconds to a key, one line to switch, and a gateway that will show you its receipts. Change the base URL and see what breaks — my bet is nothing.
More on TrustedRouter
TrustedRouter — one API, all the LLMs, provably private.
Enjoyed this essay?
Follow me for more insights on technology, startups, and the future.