We tried Headroom locally at Kelriva this week. It is a recently active open-source project built to reduce LLM token usage, and rather than read the claims, we wanted to see what it would do on a real agentic workflow.
What Headroom does
Headroom sits between your application and the model provider. It intercepts outgoing LLM calls, applies compression to the payload before it leaves, and reports exactly what it removed. The premise is straightforward: agentic systems generate a lot of token waste that has nothing to do with reasoning quality, and a proxy layer can strip that waste out without touching the output.
What we tested
We pointed our LLM calls through a local Headroom instance and ran one of our own agentic workflows against it, a mix of Claude Opus 4.6 and Claude Sonnet 5 across 39 requests. No synthetic benchmark, no cherry-picked prompt. Just a live pipeline doing what it normally does.
The results
Across the run, Headroom compressed 85.9k tokens out of 422.1k sent, a 20.4 percent reduction, and rated the compression quality High, meaning 78 percent of what it removed was identified waste rather than content that risked changing the model's output. Zero failed requests across the full run.
Most of the removed waste, 67.0k of the 85.9k tokens, came from JSON bloat. HTML noise accounted for a comparatively negligible 84 tokens. Combined with prompt caching, the dashboard reported net cache savings of $2.60 against a 77 percent hit rate (30 of 39 requests), on top of $0.125 saved directly from compression.
Why this matters
That last point, JSON bloat as the dominant waste source, is the one worth sitting with. In agentic systems, the cost is not always in the intelligence part of the workflow. A lot of it comes from repeated context, verbose tool outputs, structured payloads, and data that keeps travelling through the loop call after call, long after it has already served its purpose.
For us, this was a useful reminder that good AI engineering is not only about choosing the right model. It is also about watching the plumbing closely: what gets passed between steps, what gets repeated unnecessarily, and what a well-placed compression layer can strip out before it ever reaches the model.
Where this fits for clients
This is exactly the layer of AI engineering we look at when we build agentic workflows for clients: not just which model to call, but what actually flows through the pipeline on every call, and where the waste hides. Tools like Headroom are a sign that the practical, unglamorous side of running LLM systems is starting to get the attention it deserves. Worth watching as it matures.