AUGUR · GITHUB REPO SCORING · 0–100 · RUNS LOCALLY · NO SERVER · NO API KEY
◈ AUGUR GET STARTED →

DUE DILIGENCE FOR THE AI AGENT ECOSYSTEM

SUBSTANCE
vs LARP.

Scores GitHub repos 0–100 to tell real projects from an impressive README wrapped around empty code. Every number comes from measured data — code, commits, contributors — never a model's guess. It all runs on your machine.

GET STARTED → SEE REAL OUTPUT →

RUNS ON: your machine gh cli node ≥ 20 claude code

THE OUTPUT

1 SEARCH · 20 CANDIDATES

The evidence is the product.

Three lines from one search for "ai agent framework" — most solid to most LARP. The lines between are cut; the numbers aren't.

augur — "ai agent framework" ◉ REAL RUN 
 1. pydantic/pydantic-ai                       100
    code 13.8 MB vs README 13 KB (ratio 0.00) · last commit 1 day ago · 100+ contributors

16. UraniumCorporation/maiar-ai                 40
    code 565 KB vs README 10 KB (ratio 0.02) · last commit 12 months ago
    · zero commits in the last 12 weeks · 8 contributors · no tests

19. TransformerOptimus/SuperAGI                 21
    code 1.8 MB vs README 23 KB (ratio 0.01) · last commit 1.5 years ago
    · 17,619 stars but zero commits in the last 12 weeks · 62 contributors

HOW IT SCORES

3 SIGNALS · 1 MULTIPLIER

Three signals, one multiplier.

[ SUBSTANCE ]

40/100

Impressive README, empty code

README-to-code ratio plus absolute code volume. A giant README over 4 KB of code is the classic tell.

readmeBytes ÷ codeBytes

[ HYPE vs STAYING POWER ]

40/100

Star spike, then death

Recency of the last real commit, plus high stars paired with no recent activity — the signature of an abandoned project.

lastCommit × 52-week activity

[ HEALTH ]

20/100

Signs of real maintenance

Contributors, CI, tests, releases. Not decisive on its own — the tie-breaker between two similar repos.

contributors · CI · tests · releases

× LIVENESS MULTIPLIER

That total is then scaled by how alive the repo is (×0.35–×1.0) — a dead project is unusable, however good its code. It's why SuperAGI, abandoned 1.5 years, still scores just 21.

USE IT ANYWHERE

3 WAYS TO RUN IT

Three ways to run it.

Same scoring engine, three front doors — a Claude Code plugin, a one-line CLI, or a tool inside your own agent. All of it runs on your machine.

[ CLAUDE CODE ]

Install as a plugin

One command to add the marketplace, one to install — then ask in plain language.

/plugin marketplace add nexgl0/augur-plugin
/plugin install augur-radar@augur

Then ask: “find a python agent framework that calls tools”

[ TERMINAL · CI ]

Run the CLI directly

No agent needed — good for scripts, CI, or a quick check. Published on npm.

npx -y augur-radar "agent framework tool calling" "llm agent tools"

On npm as augur-radar · zero dependencies.

[ YOUR OWN AGENT ]

Wire it in as a tool

Any agent that can run a shell command can call augur — a Telegram or Discord bot, or your own OpenAI loop. Register one tool: the model picks keywords, the CLI returns the scored table.

import { execFileSync } from "node:child_process"

// on each tool call your agent makes, run the CLI:
const npx = process.platform === "win32" ? "npx.cmd" : "npx"
const table = execFileSync(
  npx, ["-y", "augur-radar", ...keywords]
).toString()   // = the scored table, unchanged

Scores stay deterministic — the model only chooses keywords, never the numbers.

REQUIRES: Node ≥ 20 · GitHub CLI installed and logged in (gh auth login, or a GH_TOKEN on a server). One search uses ~180–240 requests from your own 5,000/hour quota.

LIMITS

WHAT IT CAN'T DO

Where it stops.