Headroom

Headroom calls itself "the context compression layer for AI agents." It sits between your agent and the LLM provider, intercepts everything flowing into the context window — tool outputs, logs, RAG chunks, file content, conversation history — and compresses it intelligently before the model sees it. The project's headline claim: 60–95% fewer tokens with the same answers. Its own benchmarks show code search dropping from 17,765 to 1,408 tokens (92%), an SRE incident-debugging workload from 65,694 to 5,118 tokens (92%), with accuracy preserved on GSM8K and TruthfulQA. Built by Netflix senior engineer Tejas Chopra after he found himself burning through tool-heavy agent runs, it is Apache-2.0 licensed, open source, and has gathered tens of thousands of GitHub stars since its January 2026 launch (the repo now lives at headroomlabs-ai/headroom; the old chopratejas/headroom address redirects).

Key facts: three deployment modes — a Python/TypeScript library (PyPI and npm: headroom-ai), a drop-in HTTP proxy that needs zero code changes, and an MCP server exposing headroom_compress, headroom_retrieve and headroom_stats tools. A one-command wrap — headroom wrap claude|codex|cursor|aider|copilot — adds compression to popular coding agents instantly. Compression is reversible by design (CCR): originals are never deleted, and the model can pull the full data back on demand through a retrieval tool. It has been trending on X, was covered by The Register, and recently gained a dedicated OpenClaw plugin.

Plans: completely free and open source. It runs locally, so your data never leaves your machine — and the product exists to shrink your API bills, not to create a new one.

Best for: developers running tool-heavy agents (Claude Code, Cursor, Codex, Aider, Copilot) and teams watching LLM API spend. Watch out: one detailed review notes the headline compression claim was revised down for coding agents in August 2026 — treat vendor benchmarks as directional and measure on your own workloads before assuming savings. Who should skip it: non-developers, or anyone with light chat use and small contexts where compression saves little.

Related: Cursor · Claude · Mercury 3

Pros

• Sits between your agent and the LLM provider, compressing tool outputs, logs, RAG chunks, file content and conversation history before the model sees them.
• Headline claim of 60 to 95 percent fewer tokens with the same answers, with its own benchmarks showing 92 percent reductions on code search and incident debugging.
• Compression is reversible by design: originals are never deleted, and the model can pull the full data back on demand.
• Three deployment modes: a Python and TypeScript library, a drop-in HTTP proxy needing zero code changes, and an MCP server, plus a one-command wrap for popular coding agents.

Watch out for

• One detailed review notes the headline compression claim was revised down for coding agents in August 2026, so treat vendor benchmarks as directional.
• Measure on your own workloads before assuming savings; results vary by task.
• It is a developer tool for tool-heavy agents, not a consumer chatbot feature.

Pricing

Headroom is completely free and open source under the Apache 2.0 licence. It runs locally, so your data never leaves your machine, and the product exists to shrink your API bills, not to create a new one; your only cost is your own compute and the model calls you still make.

FAQs

Does Headroom delete my original data? No, compression is reversible by design: originals are never deleted, and the model can pull the full data back on demand through a retrieval tool.

How do I add Headroom to a coding agent? A one-command wrap adds compression to popular agents such as Claude Code, Cursor, Codex, Aider and Copilot, or use the library, proxy or MCP server modes.

Is Headroom free? Yes, it is Apache 2.0 licensed and open source, and it runs locally so your data never leaves your machine.

Will it cut my token bill by 95 percent? Not necessarily; the headline claim was revised down for coding agents in one detailed review, so measure on your own workloads first.