Built an MCP server that exposes SigRank's token-efficiency ranking to AI coding agents.

SigRank (signalaf.com) ranks AI operators by cascade efficiency, not raw volume. The metric:

Yield = (Cache Reads × Output) / Input²

Input is squared so re-pasting context costs quadratically; cache reuse compounds. The interesting finding: power users often score worse than casual users because they hoard context and collapse output velocity. Volume is noise; yield is signal.

The MCP server (sigrank-mcp) exposes tools for:

• Getting your current yield + rank
• Submitting signed token counts to the leaderboard
• Querying the board (all-time / 7d / 30d / 90d, platform filters)
• Head-to-head operator comparisons

Install:
npx sigrank # full local view + MCP server
npx sigrank enroll # get on the board
npx sigrank submit # sends four signed token counts (ed25519), nothing else

Privacy: open source client, runs locally, only sends four signed token counts. submit --dry-run shows the exact payload before anything leaves. The board never sees a prompt.

• MCP server: https://github.com/SunrisesIllNeverSee/sigrank-mcp
• App + leaderboard: https://github.com/SunrisesIllNeverSee/sigrank-app
• Site: https://signalaf.com

Would love feedback on the MCP tool surface — what tools would you want exposed for an efficiency-ranking layer?