SigRank MCP

Dual-mode package: interactive CLI dashboard for operators, MCP stdio server for AI clients.

npm install -g sigrank-mcp
sigrank-mcp          # launches the full tabbed TUI (TTY detected automatically)

Install from GitHub

Until npm publish is live, install directly from the repo:

# Clone + install
git clone https://github.com/SunrisesIllNeverSee/sigrank-mcp.git
cd sigrank-mcp
npm install

# Run CLI
node index.mjs                        # TUI (if TTY)
node cli.mjs board --once             # leaderboard one-shot

# Or link globally for `sigrank-mcp` command
npm link
sigrank-mcp

Repo: SunrisesIllNeverSee/sigrank-mcp Site: signalaf.com Version: 0.9.12

CLI — Commands

Default (no args) — full tabbed TUI

sigrank-mcp
sigrank-mcp tui                        # explicit — same thing
sigrank-mcp tui --platform codex       # default platform for Watch tab

Launches the full tabbed TUI. Keys: 1-4 or ← → to switch tabs · R refresh · Q quit.

board

sigrank-mcp board                      # live leaderboard, auto-refreshes every 30s
sigrank-mcp board --window 7d          # specific window: 7d · 30d · 90d · all
sigrank-mcp board --once               # print once and exit
sigrank-mcp board --window all --once

Full leaderboard view with all board metrics. Columns: # / Codename / Class / Υ Yield / SNR / Depth / Tokens / Force / Pct / 7d↕. Sorted by Υ Yield (the token cascade).

me

sigrank-mcp me                         # claude cascade across all 4 windows
sigrank-mcp me --platform codex        # different platform
sigrank-mcp me --compare               # includes raw pillar comparison inline

Single-platform cascade view with narration card.

compare

sigrank-mcp compare                    # claude — tokenpull vs ccusage vs token-dash vs tokscale
sigrank-mcp compare --platform codex   # codex verification

Raw pillar audit across all 4 sources and all 4 windows. Useful for verifying your numbers before submitting. Shows deltas between sources.

watch

sigrank-mcp watch                      # live tune meter, re-reads logs every 30s
sigrank-mcp watch --window 7d          # specific window
sigrank-mcp watch --refresh 60         # custom poll interval (seconds)

Real-time cascade updater — reads local logs on each tick and shows your current metrics.

Options

MCP Server mode

When stdout is not a TTY (i.e. piped to an AI client), sigrank-mcp starts an MCP stdio server automatically. AI clients (Claude Code, Cursor, Windsurf, etc.) use this path.

Add to .mcp.json or equivalent:

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

Or if installed globally:

{
  "mcpServers": {
    "sigrank": {
      "command": "sigrank-mcp"
    }
  }
}

MCP Tools

Cascade math

Υ Yield    = (cache_read × output) / input²
SNR        = output / (input + output)
Leverage   = cache_read / input
Velocity   = output / input
10xDEV     = log₁₀(leverage)

Math is in cascade.mjs, dependency-free. Mirrors sigrank-app/lib/ingest/bridge.ts. Canon check: MO§ES (1251211, 11296121, 128196310, 2555179769) → Υ 18436.98.

Token Pillars — sources

The dashboard pulls from multiple sources and shows them side-by-side for verification:

Codex input is estimated — Codex logs don't expose true input tokens directly. The formula:

input       = output × ioRatio         (ioRatio derived from Claude ratio, else 2.0)
cacheCreate = uncached − input         (uncached = input_tokens − cached_input_tokens)
cacheRead   = exact (from logs)

Verifier numbers (ccusage/tokscale for codex) show raw uncached input (input_tokens − cached) — a different field than the estimated input above. The discrepancy is expected and explained inline in the dashboard.

Platform adapters

All adapters are token-only (no message content, no cost fields, no credentials).

estimated=true means one or more pillars are derived, not native. The server re-scores all submitted pillars authoritatively; local preview Υ is indicative only.

Privacy

• Token-only, always. No message content is ever read, logged, or transmitted — only token counts (input, output, cache_creation, cache_read), message IDs, and timestamps.
• Local by default. tokenpull reads only ~/.claude/projects (Claude) or ~/.codex (Codex) on your device. Numbers stay on your machine unless you explicitly submit with a codename.
• Background tooling excluded. Memory plugins, observers, summarizers (e.g. claude-mem, mem0, observer-sessions) are filtered from both Claude and Codex reads. subagents/ are kept — they represent real operator work.
• No auth required. All board reads and the submit path are anonymous.

Env vars

Dev / test

node test.mjs          # 29 unit tests (no network, no fs writes)
node index.mjs         # stdio MCP server directly (pipe to MCP client)

Tests verify (29 assertions):

• rank_paste canon: MO§ES (1251211, 11296121, 128196310, 2555179769) → Υ 18436.98 · TRANSMITTER
• submit_paste preview (no codename) + POST shape (injected fetch, no live writes)
• tokenpull dedup, window slicing, 4-window pillars (mock adapter)
• tokenpull_submit all 4 windows POST, sha256 hash, ddmmyy stamp
• tokenpullCodex io_ratio conversion per-window
• Adapter registry (15 platforms) + per-adapter shape contracts
• rank_windows 4-window paste scoring, partial input, no-network
• watch_tokenpull cascade snapshot, interval_s, TODO(AUTH.WIRE) stub
• Hardening: div-by-zero guards, parsePillars warnings, fetch timeout, EXCLUDE_TOOLING regex, narrate safety

File map