One API,
multiple open models.
No meter. No cap.
Fixed monthly pricing. No token caps, no per-request bills, no watching a usage counter while your agent runs. We route each request to the right model for the job — a fast one for the easy turns, a big one when it's hard.
- Flat price from $79.99/mo
- No token caps, no 429s
- Cancel anytime, no contract
Flat rate. Fixed lanes. No surprises.
You pay for seats and concurrency, not tokens. Your agent runs in its own lane — when it bursts past the limit, we queue it. It doesn't crash, doesn't 429, doesn't show up on a bill.
Fixed concurrency, not token budgets
You buy concurrent request slots — 1, 2, or 4 depending on your tier. Run them all day. There's no token counter, no overage charge, no meter ticking while your agent works.
We queue. We don't 429.
When your agent swarm bursts past its lane limit, requests wait in the gateway queue instead of getting refused. SSE keep-alives hold the connection. Your agent sees a pause, not a crash.
Drop-in. One line.
Point your existing OpenAI SDK, LangChain, or Cursor config at
api.nocapllm.com/v1. Same request shape,
same response shape. No SDK fork, no client rewrite.
Stop picking models. Start routing.
You send one request. We decide which model runs it — a small fast model for autocomplete-grade work, a mid model for most of it, the big one for hard reasoning. Classification happens on our side, in milliseconds, based on prompt features and task type.
One POST. model="auto"
Same OpenAI SDK call you already write. No model picker, no routing config.
Prompt features, task type
We read the prompt, estimate complexity, and pick the cheapest model that handles it well. Milliseconds, not seconds.
Fast, mid, or big
Most agent traffic doesn't need the flagship. Your cost per request drops because we stop wasting it on turns that didn't need it.
Your cost drops
Most agent traffic doesn't need your most expensive model. Today every provider sends everything to the flagship and bills you for it. We route the easy turns to a cheaper model so you're not paying GPT-4 prices for a one-line tool call.
You never pick a model again
When a better open model ships next month, it appears behind the same endpoint with no code change on your side. No vendor lock-in — Z.ai will never route you to Kimi, and we will. Multi-vendor is structurally available only to an aggregator.
Open-weight. Multi-vendor. One bill.
We run the best open model available at the time, behind one endpoint. The model that serves your tier today may not be the one that serves it next quarter. Intelligence only goes up.
Starter
Early access$79.99 /mo
Model today
Qwen3.8-27B
- Concurrency
- 1 lane
- Context
- 32K
Pro
Coming soon$179.99 /mo
Model today
GLM-5.3-Flash
- Concurrency
- 2 lanes
- Context
- 256K
Max
Coming soon$389.99 /mo
Model today
GLM-5.3
- Concurrency
- 4 lanes
- Context
- 1M
The model shown for each tier changes when a better open model ships. Your endpoint and API key don't. Yearly saves 2 months. Cancel anytime.
The constraints serious engineers check.
Prompt Privacy
No prompts or completions are ever logged in our analytics (PostHog) or email tools (Loops). Gateway logs carry method, path, status, and latency — never request bodies. Prompt cache lives in GPU memory only, evicted within minutes, never written to disk, never shared across accounts.
Long-Running Sessions
A refactor that touches forty files, an overnight test loop, an incident session that runs for hours — none of it burns down a token budget, because there isn't one. Your lanes stay yours for as long as the job takes. The only thing that ends a run is your agent deciding it's finished.
The questions we actually get.
What happens when my agent goes past its concurrency limit?
Requests queue at the gateway instead of being refused. SSE keep-alives hold the connection open, so your agent sees a pause rather than a 429 and a crashed run. Nothing is dropped and nothing is billed as overage.
Does "no token limits" really mean no limits?
There is no token meter and no per-request bill. What you buy is concurrency: a fixed number of request lanes that you can run continuously. The ceiling is how many requests you can have in flight at once, not how many tokens you spend.
Which model actually serves my requests?
Each tier is backed by an open-weight model today — Starter runs Qwen3.8-27B, Pro runs GLM-5.3-Flash, and Max runs GLM-5.3. With model="auto" we classify each prompt and route it to the cheapest model that handles it well. The backing models change as better open weights ship; your endpoint and API key do not.
Do you log or train on my prompts?
No. Prompts and completions are never written to our analytics or email tools, and we do not train on your data. Gateway logs carry method, path, status, and latency only. The prompt cache lives in GPU memory, is evicted within minutes, is never written to disk, and is never shared across accounts.
Do I have to rewrite my code?
No. nocapllm speaks the OpenAI API. Point your existing OpenAI SDK, LangChain, or Cursor configuration at https://api.nocapllm.com/v1 and change nothing else — same request shape, same response shape, no SDK fork.
Can I cancel, and is there a contract?
Cancel any time, no contract and no minimum term. Monthly plans stop at the end of the period you have paid for. Paying yearly saves two months.
Seats are limited per tier.
Every tier is a fixed number of concurrency lanes on real hardware, so we open seats as capacity lands. Leave your email and we'll tell you when yours is ready.