Sigrank MCP

What SunrisesIllNeverSee/sigrank-mcp does

SunrisesIllNeverSee/sigrank-mcp measures observable token-processing patterns for AI coding operators. It works with four token pillars: input, output, cache creation, and cache reads. From those values it calculates Yield, Leverage, Velocity, signal-to-noise ratio, and 10xDEV, then assigns an operator class across 24 experience stages.

Local analysis stays on the device. The server can read supported session logs, accept pasted counts, inspect data from ccusage, token-dashboard, and tokscale, or build a portable sigrank/0.1-draft record. It also reads public operator data from SignalAF and can publish signed snapshots when the device has been enrolled.

How it works

The headline Yield metric is calculated as cache reads multiplied by output, divided by input squared. Leverage is cache reads divided by input, Velocity is output divided by input, SNR is output divided by the sum of input and output, and 10xDEV uses the base-10 logarithm of Leverage. Canonical calculations are delegated through token-cascade.

Tools can score one paste or four windows—7d, 30d, 90d, and all-time. Windowed dashboard input uses named fields because dashboard column order can differ. Diagnostic and improvement tools run locally: diagnose_cascade identifies efficiency leaks, suggest_improvements ranks possible changes, and simulate_change compares current and proposed pillar values. self_improve combines collection, diagnosis, suggestions, and simulation in one operation.

Setup and configuration

Node.js 20 or later is required. The package can run without a global installation:

npx sigrank

For MCP clients, configure a stdio entry using npx and the package name:

{
  "mcpServers": {
    "sigrank": {
      "command": "npx",
      "args": ["-y", "sigrank"]
    }
  }
}

The same package provides a terminal dashboard and MCP entry point. To publish verified data, create a connect code in SignalAF Settings, run npx sigrank enroll, inspect a dry run, and then submit. Enrollment creates a local Ed25519 keypair under ~/.sigrank-mcp/identity.json; only the public key is sent. The MCP submit_verified flow requires confirmation, and dry_run shows the payload without publishing.

Tools and capabilities

• Pull local token usage across four windows with tokenpull, or compare tokenpull, ccusage, token-dashboard, and tokscale with tokenpull_compare.
• Score pasted counts using rank_paste, rank dashboard tables with rank_windows, and create a standard compatibility record with get_sigrank_standard_record.
• Inspect public rankings with get_leaderboard, retrieve a codename profile with get_operator, and find mentors, peers, and complementary operators with discover_peers.
• Analyze and change-test cascades through diagnose_cascade, suggest_improvements, and simulate_change.
• Submit enrolled, signed local snapshots with submit_verified; preview-only paste and token-pull tools do not publish.
• Examine tokscale data through model, cost, market-share, developer, device, MCP-usage, trend, and competitive-intelligence reports.

Limitations and notes

The measurements describe token-processing patterns, not model quality, software quality, productivity, or business outcomes. Some adapters may estimate a pillar when the source does not expose it independently, and local logs can contain partial data or data gaps.

Public submission is optional. Ranked submissions include token pillars and account, device, signature, platform, version, count, and report metadata. Prompt text, response text, source code, repository contents, and local paths are excluded from the ranked payload. The optional usage proxy is separate: it handles compatible Anthropic or OpenAI traffic in memory, binds to loopback, and records token counts and metadata rather than request content.

submit_paste and tokenpull_submit are preview paths. Publishing requires the enrolled-device route through submit_verified or the SignalAF web interface. The private signing key remains on the device.