Part of
Open source · Apache 2.0
Headroom strips repeated boilerplate from tool calls, DB rows, file reads, and RAG payloads before they hit the model. Same answers, fewer tokens, cleaner traces.
Start today
Start as a wrapper, proxy, SDK call, or MCP server. Same compression pipeline, whichever surface your stack needs.
Launch Claude Code, Codex, Cursor, Aider, Copilot, or OpenClaw through Headroom and compress every tool read before it reaches the model.
pip install "headroom-ai[all]"
headroom wrap claude # Claude Code
headroom wrap codex # OpenAI Codex
headroom wrap cursor # Cursor
headroom wrap aider # AiderPoint any OpenAI-compatible, Anthropic, LiteLLM, or custom client at the local proxy and keep the app code unchanged.
headroom proxy --port 8787
ANTHROPIC_BASE_URL=http://localhost:8787 your-app
OPENAI_BASE_URL=http://localhost:8787/v1 your-appUse Headroom directly in app code when you want explicit control over messages, traces, and retrieval policy.
from headroom import compress
result = compress(messages, model="claude-opus-4-6")
response = client.messages.create(
model="claude-opus-4-6",
messages=result.messages,
)Install Headroom as an MCP server so compatible clients can call compression, retrieval, and stats tools on demand.
headroom mcp install
# Exposes tools:
# headroom_compress
# headroom_retrieve
# headroom_statsEnterprise & Scale
Deploy Headroom in your VPC, customize compression pipelines for proprietary formats, or get dedicated SLA support for high-volume agent workloads.