COMING SOON
Every time your AI agent pauses, it starts reading from page one.
And you pay for it. Again.
The problem
An agent doesn’t answer once and stop. It works in steps — searching, writing code, checking, fixing.
Between every step it waits. And while it waits, it loses its place.
So it reads the whole conversation again. Every step, another reading, another bill.
Measured by Microsoft, June 2026
90%of the conversation is still in memory while an agent is mid-task
55%after it pauses and comes back
172smedian gap between turns, with the conversation’s memory sitting idle the whole time
From GitHub Copilot production traces: 3.2 million users, 761 million model calls, 95 trillion tokens. We haven’t found the equivalent published for an Indian workload. We intend to.
Read the paper →How we solve it
-
We see the pause coming
A tool call means a wait is starting. That’s our signal.
-
We hold its place
Instead of losing the conversation, we set it aside.
-
We hand it back
It picks up where it stopped. No re-reading. No second bill.
Pricing
Priced in rupees, per token.
Rate card after our first benchmark. We would rather leave it blank than invent it.
- Per token Tiered by volume and by how long a request can wait Coming soon
- Per task Priced on cache-hit improvement, not raw tokens Coming soon
- Per call For Indic-heavy workloads, where tokens misrepresent cost Coming soon
Why India
Cheaper machines. Billed in rupees, not dollars. And Indian languages cost four times more to process than English — so we price them properly.
We're building, not selling.
If you're running agents in India and paying too much for them, we'd like to talk. Not to sell you anything — to understand what actually hurts.
Built on vLLM, SGLang, LMCache and TensorRT-LLM.
Email usModels we plan to serve.
Open weights only. Nothing we serve is ours, and nothing we serve is locked to us.
Need one that isn’t here? Tell us and we’ll look at serving it.
