Private preview
Streamproof is a drop-in OpenAI-compatible endpoint that fixes two specific things when you run LLMs for clients: silent empty replies from cheap models in Continue and Cline, and a runaway bill from one shared upstream key.
New and small — you’d be one of the first users. Prices are in EUR, excluding VAT.
We sell against these two things and nothing else. If neither is your problem, the free stream fix is still worth having.
“I gave the agency client the same key I use. Now I can’t tell whose usage is whose, and I can’t stop them from burning €400 in a weekend. I found out from the invoice.”
“Continue just returns nothing. No error, no reply. I spent an afternoon thinking it was my prompt.”
Three steps, done with you in about a day for the Standard setup.
Stand up the proxy in your own Vercel project and point it at your upstream base URL and key. The key stays an environment variable in your project — we never receive it.
Every client key gets a monthly USD budget and an RPM limit. Keys are stored only as SHA-256 hashes; the plaintext is shown once, when it is issued.
Continue and Cline keep working with no config changes beyond the base URL. Over budget returns 402, over rate returns 429, a bad key returns 401. It fails closed.
Strips rejected request fields (like min_p), drops empty or reasoning-only chunks, removes per-chunk usage, and preserves tool_calls. The empty-reply bug is fixed, not documented.
Over a client’s monthly budget the request is refused with 402 instead of billed. A retry loop becomes an error at the limit, not a surprise invoice.
Your Vercel project, your upstream keys, your spend. There is no shared multi-tenant service your traffic passes through, and no lock-in.
See spend, requests and budgets per key, then export the raw events as CSV. Built for people who bill clients for this.
No runtime dependencies in the proxy, Web Request/Response APIs, pure functions for the repair logic. You can read the whole thing.
No uptime SLA, no 24/7 on-call, no invented customer count. It is a small product with one maintainer, and it says so.
The stream-repair approach builds on the MIT-licensed Continue-NIM-Proxy project. Credit to its author.
A one-time setup, plus an optional monthly managed plan. Your own model/API spend is never included — you pay your upstream provider directly.
€300one-time setup
One project, one provider, a handful of clients.
€600one-time setup
More clients, more than one upstream, a dashboard to show.
€29/month
1 workspace, up to 5 client keys.
€59/month
1–3 workspaces, up to 20 client keys.
€99/month
Unlimited workspaces, 30+ client keys.
Setup is a fixed price, paid before work starts — no hourly billing. The managed plan is optional and you can cancel in one email.
In your Vercel project. You own the deployment, the domain and the environment variables. We don’t run a shared service that your traffic passes through.
Your upstream provider key is an environment variable in your Vercel project. We never receive it. Client keys are stored only as SHA-256 hashes — the plaintext exists once, when it is issued.
Every client key has a monthly USD budget and an RPM limit. Over budget returns 402, over rate returns 429, an unknown key returns 401. It fails closed.
The proxy forwards requests and bills on token counts; it does not need to retain prompt or completion content to do that. Your upstream provider processes the request under their terms, same as today.
If you already run one and your clients’ streams work, keep it — you don’t need us. What those don’t give you out of the box is the stream repair for the specific clients you use, plus a per-client hard cap you can set in two minutes.
The deployment is yours. You keep the code and the running Vercel project; you only lose monitoring and updates. There’s no lock-in because there’s no mechanism for one.
Honest answer: it’s a small product with one person behind it. It fails closed — over budget returns 402, not an unbounded bill. Support is best-effort with a stated response window; there’s no 24/7 on-call and we won’t pretend otherwise.
No. You pay your upstream provider directly for tokens, at their prices. Streamproof charges for the setup and (optionally) the managed plan — never a markup on your model spend.
The stream repair is useful on its own. If you later hit the shared-key and budget problem with a client, the setup is here.