sigrank-mcp

AI operator leaderboard that ranks how efficiently you use AI coding tools, not the models themselves. 15 MCP tools, TUI, ed25519-signed submissions, privacy-first.

GitHub

Documentation

SigRank MCP

🏆 SigRank is live: signalaf.com — the leaderboard for how efficiently you use AI, not how much. See your projected rank in 60 seconds at signalaf.com/score. Token counts only. Never your prompts.

SigRank — the new standard in AI evaluation & benchmarks

The yield cascade + live leaderboard as MCP tools any agent can call.

For all builders, burners and 10xers.

Quickstart
Install from GitHub
Install via Smithery
Commands
MCP Server mode
Cascade math
Token Pillars
Platform adapters
Privacy
Env vars
Dev / test
File map
Contributing
License

The board	Your operator profile

Every operator ranked by Υ Yield — the architecture of the cascade, not raw spend	Cascade layer, class, and fingerprint — derived from four token counts

Run sigrank enroll then sigrank submit to get ranked and claim your public profile at signalaf.com.

Quickstart — 3 steps to the board

# 1. Install (pulls ccusage + tokscale + tokendash automatically — no separate installs)
npm install -g sigrank

# 2. Sign in (paste a connect code from signalaf.com → Settings → New key)
sigrank enroll

# 3. Submit your cascade to the board
sigrank submit

# (cautious? see exactly what would be sent — four counts + a signature — sending nothing)
sigrank submit --dry-run

That's it. sigrank reads your local AI session logs on-device, derives your token cascade (Υ Yield, Leverage, Velocity, 10xDEV), and publishes to signalaf.com. No paste, no transcript content — only the four token counts leave your machine.

Or just explore without signing in:

sigrank          # launches the full tabbed TUI (dashboard, compare, board, watch)
npx sigrank board --once    # print the live leaderboard once

Install from GitHub

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` command
npm link
sigrank

Repo: SunrisesIllNeverSee/sigrank-mcp Site: signalaf.com npm: sigrank Smithery: smithery.ai/servers/burnmydays/sigrank-mcp Glama: glama.ai/mcp/servers/SunrisesIllNeverSee/sigrank-mcp

Install via Smithery

SigRank is available on Smithery as a stdio MCP bundle — one-click install for Claude Desktop, Cursor, and other MCP clients.

Smithery CLI

# Install Smithery CLI
npm install -g smithery

# Connect to SigRank (downloads the MCPB bundle locally)
smithery mcp add burnmydays/sigrank-mcp --id sigrank

# List available tools
smithery tool list sigrank

# Call a tool
smithery tool call sigrank get_leaderboard '{}'
smithery tool call sigrank rank_paste '{"text": "1000000 500000 50000 800000"}'

Claude Desktop (via Smithery)

Go to smithery.ai/servers/burnmydays/sigrank-mcp
Click Install
Smithery handles the rest — no manual config editing

Commands

⊙ SigRank CLI  v0.16.0

Default (no args)
  sigrank              unified dashboard: cascade + token pillars + board

Commands
  enroll                   sign in: paste a connect code (get one at signalaf.com → Settings)
  submit                   publish your verified runs to the board (sign in first)
  board                    live leaderboard (refreshes every 30s)
  board --window 7d        board for a specific window (7d, 30d, 90d, all)
  board --once             print once and exit
  compare                  raw pillar audit: tokenpull vs ccusage vs token-dash vs tokscale
  compare --platform codex compare for a specific platform
  tui                      full tabbed TUI: Dashboard / Trends / Compare / Board / Watch / Connect
  tui --platform codex     TUI with a different default platform
  watch                    live tune meter — ALL active platforms × all windows, every 30s
  watch --platform codex   watch only one platform (optional filter)
  watch --window 7d        watch only one window (optional filter)

Options
  --window    7d · 30d · 90d · all  (default: 30d for board; all windows for watch)
  --platform  claude · codex · amp · gemini · opencode · goose · …
  --refresh   poll interval in seconds (default: 30)
  --once      print once and exit (board only)

For AI clients (not typeable)
  In a piped/non-TTY context, sigrank is an MCP stdio server.
  AI clients (Claude, Cursor, …) call its tools automatically — these are
  NOT shell commands. Humans use the commands above.

Examples
  sigrank                        # unified dashboard
  sigrank board                  # live leaderboard
  sigrank compare                # pillar audit (claude)
  sigrank compare --platform codex
  sigrank watch --window 7d --refresh 60
  sigrank board --window all --once

The TUI is the whole app

Launch it and sign in inside it:

npx sigrank

Six tabs. Keys: 1-6 or ← → to switch · R refresh · Q quit.

Tab	Key	Content
Dashboard	`1`	Cascade table (all platforms × windows + combined) · Υ sparklines · token composition bars · mini board
Trends	`2`	Every metric across windows — sub-views: You / Platform / Field
Compare	`3`	4-source pillar audit (tokenpull vs ccusage vs token-dash vs tokscale) · delta % · cascade metrics per source · cache read bar chart
Board	`4`	Full leaderboard with all fields · `[W]` cycles window (7d/30d/90d/all)
Watch	`5`	In-TUI landing panel · `[Enter]` launches the live watcher (big numbers + pillar bars + Υ trend, auto-refreshes 30s)
Connect	`6`	Sign in / switch device — paste a connect code from signalaf.com → Settings. Then `[S]` submits.

Sign in + submit

sigrank enroll          # sign in: paste a connect code (get one at signalaf.com → Settings)
sigrank submit          # publish your verified runs to the board (sign in first)
sigrank submit --dry-run  # inspect the exact signed payload without sending anything

Or do it inside the TUI on the Connect tab (6), then press [S] to submit.

MCP Server mode

When stdout is not a TTY (i.e. piped to an AI client), sigrank 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"]
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "sigrank": {
      "command": "sigrank"
    }
  }
}

MCP Tools

Tool	Args	What
`rank_paste(text)`	`{input, output, cacheCreate, cacheRead}` JSON or 4 whitespace-delimited numbers	Scores token pillars → Υ Yield / SNR / Leverage / Velocity / 10xDEV / Class + prose narration card
`get_leaderboard()`	`{window?}`	Live board from signalaf.com — sorted by Υ Yield
`get_operator(codename)`	`{codename}`	One operator's live profile
`submit_paste(text, codename)`	`{text, codename?}`	Rank locally then POST to board. Omit codename for preview-only
`tokenpull(platform?)`	`{platform?}`	On-device local reader: scans local logs → 4-window cascade. Zero paste, token-only
`tokenpull_submit(codename, window?)`	`{codename?, window?}`	`tokenpull` → publish to board. Omit codename for preview
`tokenpull_compare(platform?)`	`{platform?}`	All four sources side-by-side: tokenpull + ccusage + token-dash + tokscale. Returns pillars, cascade metrics, and delta % vs tokenpull per window
`rank_windows`	`{platform?, window?}`	Multi-window cascade from local logs
`watch_tokenpull`	`{platform?, interval_s?}`	One cascade snapshot per call (interval_s advisory)
`submit_verified`	`{window?, platform?, dry_run?}`	THE ranked path: builds + ed25519-signs Schema 1.0 snapshots and POSTs them. `platform:'multi'` sums all active platforms. `dry_run:true` returns the exact payload unsent
`enroll`	`{code, device_label?}`	Bind this device with a connect code from signalaf.com → Settings
`diagnose_cascade`	`{text?}`	Diagnoses where your token cascade is leaking efficiency — ranked findings with severity + estimated Υ impact
`simulate_change`	`{text?, changes}`	Prescriptive "what if" — test proposed pillar changes and see the exact Υ delta + class change before committing
`suggest_improvements`	`{text?}`	Generates ranked, simulated improvement suggestions — tests strategies and returns them sorted by Υ yield impact
`self_improve`	`{text?}`	One-click optimize: diagnoses, suggests, and simulates the best change in a single call

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:

Source	What	Platform
`tokenpull`	On-device JSONL scanner (canon source)	claude, codex, amp, …
`ccusage`	`ccusage <platform> daily --json` CLI (bundled)	claude, codex
`token-dashboard`	`~/.claude/token-dashboard.db` SQLite (bundled)	claude only
`tokscale`	`tokscale models --json` CLI (bundled, falls back to `~/tokscale_report.json`)	claude, codex

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).

Platform	Path	Notes
Claude Code	✅ `~/.claude/projects`	Native; dedup by `(session_id, message_id)`; subagents included
Codex	✅ `~/.codex/sessions`	Estimated input via `io_ratio`; verified vs ccusage
Amp	✅ `~/.local/share/amp/threads`	Full 4-pillar; per-message
Kimi	✅ `~/.kimi/sessions`	Full 4-pillar; `StatusUpdate` lines only
pi-agent	✅ `~/.pi/agent/sessions`	Full 4-pillar; per-message JSONL
OpenClaw	✅ `~/.openclaw`	Full 4-pillar; per-message JSONL
Droid	✅ `~/.factory/sessions/*.settings.json`	Full 4-pillar; thinking→output
Codebuff	✅ `~/.config/manicode`	Full 4-pillar; `chat-messages.json`
Hermes	✅ `~/.hermes/state.db`	Full 4-pillar; SQLite; reasoning→output
Kilo	✅ `~/.local/share/kilo/kilo.db`	Full 4-pillar; SQLite
Qwen	✅ `~/.qwen/projects`	`cacheCreate=0` estimated; thought→output
Goose	✅ `~/.local/share/goose/sessions/sessions.db`	`cacheCreate=cacheRead=0` estimated; SQLite
Gemini CLI	✅ `~/.gemini/tmp`	`cacheCreate=0` estimated; cache extracted from input field
GitHub Copilot CLI	✅ `~/.copilot/otel`	OTel JSONL; requires `COPILOT_OTEL_ENABLED=true`
OpenCode	⚠️ `~/.local/share/opencode`	Raw token counts not persisted in log format
Cursor	🔜	Chat log path TBD
Windsurf	🔜	Session logs at `~/.codeium/windsurf/`

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.
Board reads are anonymous. No account needed to browse, compare, or watch.
Ranked submissions are signed, not trusted. sigrank submit requires a one-time enroll (device-bound ed25519 key — the private key never leaves your machine). Verify what's sent with sigrank submit --dry-run: the payload is four token counts, ratios, and a signature.

Env vars

Var	Default	Description
`SIGRANK_API_BASE`	`https://signalaf.com`	Override the board host
`SIGRANK_FETCH_TIMEOUT`	`10000`	Board API fetch timeout (ms)

Dev / test

node test.mjs          # 13 test groups, 200 assertions (no network, no fs writes)
node sign.test.mjs     # ed25519 signing + canon parity
node index.mjs         # stdio MCP server directly (pipe to MCP client)

Tests verify (13 groups, 200 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, submit path
enroll posts identity (public key only), maps 201 enrolled + 410 code_invalid
submit_verified signs Schema 1.0, server-verifiable
simulate_change relative + absolute deltas, quadratic penalty, JSON input
Hardening: div-by-zero guards, parsePillars warnings, fetch timeout, EXCLUDE_TOOLING regex, narrate safety
sign.test.mjs ed25519 round-trip + canonical 926-byte payload parity

File map

File	Responsibility
`index.mjs`	Entry point — TTY detection, routes to CLI or MCP server
`cli.mjs`	CLI commands: board, compare, watch, enroll, submit, help
`tui.mjs`	Full tabbed TUI: Dashboard / Trends / Compare / Board / Watch / Connect
`cascade.mjs`	Pure cascade math (Υ, SNR, leverage, velocity, 10xDEV, class)
`tokenpull.mjs`	On-device log scanner — Claude, Codex, multi-platform
`adapters.mjs`	Platform adapter registry (15+ platforms)
`tools.mjs`	MCP tool table + dispatcher
`connect.mjs`	Connect-code enrollment + device identity
`keystore.mjs`	Local key management (paste-keys, not API keys)
`submit.mjs`	Verified submit flow (signs + POSTs to board)
`sign.mjs`	Schema 1.0 signing (X-Agent-Signature)
`narrate.mjs`	Deterministic prose narration card
`preflight.mjs`	Plausibility checks (Benford, bounds, anomaly detection)
`test.mjs`	Unit tests (no external deps)
`sign.test.mjs`	ed25519 signing + canon parity test

Contributing

Contributions welcome. SigRank MCP is built in the open.

Contributing guide
Security policy
Report bugs via GitHub Issues
PRs: fork → branch → node test.mjs passes → open PR against main

License

MIT — see LICENSE.