Updated
Does a bigger context window make rereading docs cheaper?
No. A bigger context window raises how much you can load, not how efficiently. You still pay for every token, every session, every agent, so rereading docs is as expensive per token, you just do more of it. A bigger window isn't a more current one either. Serving one canonical answer per query is about 60% fewer tokens per lookup.
Why doesn't a larger window lower the cost?
Because pricing is per token, not per window. A 1M-token window doesn't discount the 20k tokens you load into it — it just lets you load more. If your agent rereads the same runbooks and READMEs each session to find which is current, a bigger window means it can reread more of them, not more cheaply. The cost compounds across sessions, agents, and teammates regardless of window size.
Doesn't more context at least mean better answers?
Not on its own. A bigger window can hold the stale copy just as easily as the canonical one, so it does nothing for correctness — and flooding the window with near-duplicate files can dilute the relevant part and make answers worse — models attend least reliably to information buried in the middle of a long context, and accuracy drops as the context grows, even on models built for long context (Liu et al., Lost in the Middle, TACL 2024). More context is not the same as the right, current context.
What lowers the cost for real?
Loading less, but the right thing. Route each query to the single canonical doc, read the section that answers, and skip stale and duplicate copies. trovex does exactly this as an open-source, local-first MCP server: one current answer per query with a freshness marker. It shows you the tokens it saved, independent of how big your model's window is.
FAQ
If the window is huge, why does cost still climb?
Because a bigger window changes how much the agent can hold, not how much it needs to read. It still reopens candidate docs each session to find the current one and still pays for every token it reads. Capacity is not the bottleneck; knowing which doc is current is.
Can I just paste all the docs into the window?
You can, and you pay for all of them on every turn, plus the model has to find the answer inside the pile. Serving the one canonical doc per query is cheaper and sharper than dumping everything into a large window.
What actually reduces the rereading?
A freshness signal and routing, not size. trovex marks each doc canonical, stale, or duplicate and returns only the current one for a query, so the agent reads one section instead of re-scanning candidates. About 60% fewer tokens per lookup on our own repo.
Load less, the right thing.
Serve your agents one canonical answer per query — whatever your window size.
uv tool install trovex
Open source. No cloud, no API keys. Your docs never leave your machine.