Updated
How do I manage context for my coding agents?
Managing context for coding agents is controlling what they read — not hoping a bigger window saves you. Give each query one canonical doc with a freshness marker, read at the section level, and let every agent share one write-back path so stale and duplicate copies stop being re-read. About 60% fewer tokens per lookup.
What "managing context" actually means
It's not prompt-crafting and it's not buying a longer window. The thing you manage is the read: which doc an agent opens for a question, whether that doc is current, how much of it loads, and whether the next agent re-derives the same answer. Left unmanaged, every session re-scans the repo to rebuild what was already known — and you pay for it each time. Managed, each lookup is cheap and final.
The four things you're actually managing
| What you manage | The lever |
|---|---|
| Which doc answers a question | Route each query to one canonical doc |
| Whether it's the current one | Freshness markers skip stale copies |
| How much loads per answer | Section-level reads, not whole files |
| What the rest of the fleet sees | Write-back so agents share one answer |
Why the usual moves don't manage it
A bigger context window raises the ceiling, not the efficiency — the agent still rereads to find which copy is current. A growing AGENTS.md / CLAUDE.md helps at five docs and hurts at fifty: it loads in full every turn whether or not it's relevant. Both leave the read unmanaged. Managing context means the agent asks one place and gets one current section back.
How trovex manages it
trovex is an open-source, local-first MCP server built around these four levers: one canonical answer per query, freshness markers, section-level reads, and a shared write path the whole fleet reads and writes through. It runs on your machine (SQLite + ONNX, no cloud or keys) and keeps a savings dashboard, so the reduction is a number you can check on your own repo — not a claim you take on faith.
Manage the read, not the window.
Index your repo and let trovex serve one current answer per query.
uv tool install git+https://github.com/TsukumoHQ/trovex
Open source. No cloud, no API keys. Your docs never leave your machine.