Anyone who has watched Claude Code or a similar coding agent burn through tokens reading files knows the problem: a single grep or file read can dump thousands of irrelevant lines into the context window (the amount of text the AI model can "see" at once). Most of that content is noise the model never needed.
Context Gateway is a new open-source proxy designed to fix this. It sits between your coding agent and the LLM API, intercepting tool outputs and compressing them before they enter the context window. The result is less token waste, lower API costs, and potentially better model performance since the AI isn't sorting through pages of irrelevant code to find the three lines that matter.
The project supports agents like Claude Code and OpenClaw out of the box. Because it works as a proxy rather than a plugin, it doesn't require changes to the agent itself. You point your agent at the gateway instead of directly at the API, and the compression happens transparently.
This addresses a real pain point. Context management is one of the weakest parts of current AI coding agents. They tend to be greedy with reads, pulling in entire files when they need a single function. That bloats costs and can push important earlier context out of the window as newer (but less relevant) content takes its place. A compression layer that strips the noise before it reaches the model is a practical solution, even if it's treating the symptom rather than fixing agent behavior directly.
The project is available on GitHub under open-source licensing. It's early-stage, so expect rough edges, but the core idea is sound enough that something like this will likely become standard infrastructure for agent-heavy workflows.