Why do AI agents burn so many tokens on tool calls?

Every tool output (search results, DB queries, file listings, logs) serializes its full JSON response into the context window. By turn 10 the agent carries 100k+ tokens of prior tool responses on every API call even when only 500 tokens are relevant.

What is headroom and how does it work?

headroom is an open-source context compression layer (library + proxy + MCP server) that compresses tool outputs before they reach the LLM. It achieves 60-95% token reduction using reversible compression; the original content can be retrieved on demand via headroom_retrieve.

How fast did headroom grow?

headroom launched on GitHub in January 2026 and reached 26k stars and 1.7k forks by June 2026, making it one of the fastest-growing MCP ecosystem tools in 2026.

Why is there no built-in compression in agent SDKs?

OpenAI Agents SDK, LangChain, and Claude tool use APIs all pass tool responses directly into context without compression. The problem was under-recognized until agents became long-running and multi-turn, which pushed token costs into the $10-250/day range per session.

What context window percentage do tool schemas alone consume?

One team audited at 143,000 of 200,000 tokens (72% of their context) spent on MCP tool definitions alone before any conversation content. This is the tool schema bloat problem, separate from tool output bloat.

← Back to dashboard

clawsmith.com/signal/agent-tool-output-bloat-context-cost-explosion

📈 TrendsUnderservedToolLive

AI agent tool outputs bloat context window and burn tokens with no compression layer

Each tool call (search result, DB query, file listing) serializes its full JSON response into the context. By turn 10 the agent is paying for 100k+ tokens on every call. headroom reached 26k GitHub stars in 5 months solving this with 60-95% compression. There is no built-in compression in any major agent SDK or MCP client.

Score Breakdown

GitHub

27,833

Issues

Social Proof 3 sources

headroom: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM

gh:chopratejas · 1/7/2026

27,833 GH

[FEATURE] headroom wrap opencode — CLI wrapper for OpenCode (73 reactions)

2/1/2026

78 HN

Show HN: Headroom – Reversible context compression for LLMs (~60% cost reduction)

3/1/2026

Existing Solutions 1 competitor

headroom26k GitHub stars, Jan 2026 launch. Early-stage, no enterprise support

Reversible context compression layer: library, proxy, MCP server. 60-95% token reduction on tool outputs

Gap Assessment

UnderservedExisting solutions leave gaps

headroom is early-stage (Jan 2026) and covers one approach; no first-party compression layer exists in OpenAI Agents SDK, LangChain, or Claude tool use

Frequently Asked Questions

Virality Score

27,913

across 0 platforms

Details

Signaltrend

EcosystemTool

Sources3

Platforms0

Updated2h ago

Trend→ stable

Top ideas

All ideas →

0A web app that shows indie mobile developers exactly why their app is or isn't getting organic App Store discovery 0A web app that lets API and SaaS builders charge sub-dollar per-request fees without hitting Stripe minimums 0A mobile app that replaces Mint with free bank sync, automatic categorization, and spending alerts