Token-efficiency proxy for Codex/OpenAI Responses API: recoverable storage, MCP server, opt-in visual compaction, effort routing
View on GitHub ↗TokenSkein is a local context-economy gateway for GPT clients and Codex. It combines the useful ideas behind pxpipe, Headroom, RTK, and Caveman-style output discipline without pretending that every optimization helps every request.
The name combines token with skein: schemas, recovered text, image pages, shell output, style, and reasoning are separate strands managed as one measured bundle.
- proxy OpenAI Responses API requests; - compact tool schemas and older tool outputs; - archive every compacted output locally so exact text remains retrievable; - expose compression, retrieval, and statistics over MCP; - route reasoning effort from task complexity; - inject a concise-output policy with safety exceptions; - optionally encode very dense text as PNG pages when an explicit economic gate passes; - optionally place a safe output filter after RTK for recognized read-only shell commands.
Status: working experimental MVP. Text compaction, storage, MCP, routing, proxying, and the shell adapter are implemented and tested. The vision lane is implemented but disabled by default until model-specific A/B evaluation establishes its real cost and quality.
Context cost has several different causes, so one trick is not enough:
- repeated schemas waste input tokens; - large logs and JSON consume context even when only a few lines matter; - verbose responses consume output tokens; - easy tasks do not always need high reasoning effort; - dense visual text can sometimes cost fewer model tokens than raw text, but can also damage exactness or cost more; - lossy compression becomes dangerou
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/flaviomartil--token-skein.html)
runs anywhere. uses anything
pueschel88/Tradingview-MCPTradingview MCP server for controlling Tradingview Desktop from Claude Code via Chrome DevTools Protocol
0xsline/OpenChatCutLocal-first conversational AI video editor with a professional multi-track timeline, Agent Skills, MCP integration, and Remotion-powered rendering.
EndymionLee/PilotBrowseMCPA browser runtime that lets AI agents control your real Chrome browser via MCP. Agents can explore websites, generate operation manuals, and reuse them to save tokens. AI操控你的真实浏览器。
herry2059/project-os-for-codexOpen-source control plane for Codex projects: Git-backed context, visible agent progress, scoped MCP access, resumable work, and safe handoffs.
JesusRS1/stock-trade-finance-apiStock Trade MCP server in Tiingo with stock prices, news, forex, and fundamentals programmatically via the MCP
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.