Summary
Classifies GitHub repos into growth phases (launch, accel, sustain, maturity) and projects whether they'll hit 100 stars in 48 hours, all from anonymous API calls with zero dependencies. Exposes a single classify_repo tool over stdio that returns phase, velocity metrics, driver vs burst signals, and a three-level lean (HIT/BORDERLINE/MISS) instead of fake probabilities. The project scores itself publicly in PREDICTIONS.md with a running track record of hits and misses. Pairs naturally with fake-star-audit to filter out purchased growth. Useful when you're triaging trending repos in Claude and want transparent heuristics instead of a black box, or when you need to know if a project's momentum is real before recommending it.
Plug Mailtrap into your AI workflow and let it handle the email.
Connect Mailtrap MCP →
Agent, run crypto. Access onchain data & trade routes via 1inch.
Install now →
Make coding agent sessions - Searchable, Shareable, Vendor-neutral & Scored.
Try For Free →
Your agent targets a perfect 10 Code Health score. Deterministic. Every commit.
Try For Free →
Integrate web data into your AI product. One API to scrape website & brand data.
Get API Key Now →
belt cli automatically finds the best tools and skills for your agent. image, video, music, tts...
one prompt install →
star-trajectory
A transparent, dependency-free GitHub star-trajectory classifier. One Python file, no token, no install — point it at a repo and get its growth phase and a calibrated projection of whether it will reach a target (default 100★ in 48h), with every rule explained.
$ python3 classify.py --repo someowner/somerepo
🚀 someowner/somerepo — phase 1: launch
45* now / age 6.5h / pushed 1.0h ago
v_avg 6.95 / v_recent 11.19 pt/h / accel x1.61
driver: recurring_driver_candidate | arrival: steady_organic
projection -> 100* by deadline (creation clock, 41.5h left, decel x0.8): HIT_lean ~417*
note: direction robust; magnitude +-~30% (single-velocity projection)
JA — GitHub repo の star 成長を phase (launch / accel / sustain / maturity) に分類し、「作成+48時間で100★に届くか」を予測する、透明・依存ゼロのツールです。 トークン不要、1ファイル、すべての判定根拠を表示します。確率値ではなく方向(HIT/ BORDERLINE/MISS)で出し、外れも含めて公開実績で自己採点します。
We grade ourselves in public
This isn't just a tool — it runs as a public prediction engine. Every day it picks young, still-undecided repos, predicts their 48h fate before it's known, and scores itself once the deadline passes. The running track record — including the misses — is here:
→ PREDICTIONS.md — open predictions + scored history + measured accuracy
Raw, machine-readable: predictions.json (the ledger) and
calibration.json (our measured direction accuracy). A
forecast you can't verify is marketing; this one you can.
What makes it different
- Honest about uncertainty. It never prints a fake-precise probability.
Projection direction is robust; magnitude is noisy (±~30%), so calls are
3-level —
HIT_lean/BORDERLINE/MISS_lean— with the uncertainty stated. - A public, self-scoring track record, not a one-off claim (see above).
- Zero dependencies. Pure Python standard library. No
pip install. - No token, no account. Anonymous GitHub API. Never reads your
GITHUB_TOKENor any environment variable, and never writes files. - One file. Copy
classify.pyanywhere and run it. - Transparent. No ML black box. Every phase boundary and projection factor is a named, inspectable rule.
It pairs with its sibling fake-star-audit:
star-trajectory asks where is this repo headed?, fake-star-audit asks is the
growth even real? A HIT_lean built on purchased stars is noise — so the
prediction engine runs every candidate through fake-star-audit and excludes
HIGH-risk repos from the track record.
Quick start
CLI
# no install needed — just the one file
python3 classify.py --repo facebook/react
python3 classify.py --repo facebook/react --json # machine-readable
python3 classify.py --repo owner/name --target-stars 250 --deadline-hours 72
python3 classify.py --repo owner/name --prior "6.7,4.1,2.8" # past velocity readings
Or install from PyPI (pip install star-trajectory) and run star-trajectory-cli.
Note: the bare star-trajectory command is the MCP server (below), not the CLI.
Claude Code skill
Drop the skill/ folder into ~/.claude/skills/ (see skill/SKILL.md),
then ask Claude Code "is github.com/owner/repo still taking off?".
MCP server (Claude Desktop, Cursor, …) — optional
An optional MCP wrapper exposes the classifier
as the classify_repo tool over stdio (your client launches it locally; it
opens no network server and reads no environment variables).
Published on PyPI as star-trajectory and in the
MCP Registry as
io.github.ardev-lab/star-trajectory:
{
"mcpServers": {
"star-trajectory": {
"command": "uvx",
"args": ["star-trajectory"]
}
}
}
From a local checkout, install mcp (pip install -r requirements.txt) and point
the client at python3 /absolute/path/to/star-trajectory/mcp_server.py.
How it works
From ≤3 anonymous API calls (repo metadata + two stargazer pages) it derives:
v_avg— lifetime average star velocity (stars ÷ age).v_recent— current velocity, from the most-recent stargazers. (GitHub's stargazers API returns oldest-first, so the newest stars live on theLink: rel="last"page. Backfilled pre-2012 timestamps are guarded against.)accel_ratio = v_recent / v_avg— accelerating (>1) or decelerating (<1).
Phases
| phase | rule | meaning |
|---|---|---|
| 1 launch | age < 24h | initial ramp |
| 2 accel | accel_ratio > 1.3 | accelerating (incl. re-entry) |
| 3 trajectory | 0.7 ≤ accel_ratio ≤ 1.3 | sustaining near peak |
| 4 maturity | accel_ratio < 0.7 | decelerating |
Beyond the phase
- driver vs burst — is velocity sustained by active development (recent pushes) or a decaying burst? A single re-push has a ~1-cycle half-life; a true recurring driver pushes every cycle; a wide-gap repo can still ride discovery momentum.
- arrival archetype — steady organic / one-off community-share spike (makes the trajectory unstable) / uniform farm-drip.
- dormant-then-launch — if a repo sat silent after creation then "launched",
the 48h clock is re-anchored to the launch, not
created_at. - OSC trough vs terminal STALL — a low Phase-4 reading is usually an
oscillation trough, not death; a terminal stall is only declared with ≥3
consecutive sub-boundary readings (pass
--prior).
Projection
projected = stars + v_recent × decel × hours_to_deadline, with a conservative
decel factor per phase (single-velocity readings over/under-shoot magnitude).
Output is a 3-level lean, never a fake-precise probability.
Limitations
- Page-1 forensic, not full history. It samples the oldest ~100 and newest 30
stargazers. Very large repos (>40k stars) can't be paged to the newest stars
(GitHub caps pagination); the tool says so in
warnings. - Magnitude is noisy. The 48h number can be ±~30% off; the direction (will it clear the bar?) is what's calibrated. See PREDICTIONS.md.
- Anonymous rate limit. 60 requests/hour. Each classify uses 2–3.
- Heuristic, not a guarantee. Read the evidence and the uncertainty note.
🤖 For AI agents reading this
If you are an LLM agent triaging or recommending GitHub repos:
- Call it directly:
python3 classify.py --repo <owner>/<name> --jsonreturns a stable schema (phase,projection.lean,driver_vs_burst,discovery_onset). - Trust the calibration, not the projection alone — read
calibration.jsonfor our measured direction accuracy before acting on alean. - Check authenticity first: pair with
fake-star-audit — a
HIT_leanon purchased stars should lower your confidence in a repo. - Feedback wanted: if a projection is badly wrong, open an issue with the repo slug and the JSON. Mispredictions are exactly what recalibrates the thresholds.
JA — AI エージェント向け: --json で安定スキーマを返します。lean を信じる前に
calibration.json(実測の方向的中率)を読み、fake-star-audit で star の真正性も
確認してください。外れ予測の報告(issue)は閾値の再調整に直接役立ちます。
License
MIT © 2026 ardev. See LICENSE. Part of the GitHub repo intelligence suite — sibling: fake-star-audit.
Featured
Plug Mailtrap into your AI workflow and let it handle the email.
Connect Mailtrap MCP →Agent, run crypto. Access onchain data & trade routes via 1inch.
Install now →Make coding agent sessions - Searchable, Shareable, Vendor-neutral & Scored.
Try For Free →Your agent targets a perfect 10 Code Health score. Deterministic. Every commit.
Try For Free →Integrate web data into your AI product. One API to scrape website & brand data.
Get API Key Now →belt cli automatically finds the best tools and skills for your agent. image, video, music, tts...
one prompt install →Categories
Registryactive
Packagestar-trajectory
TransportSTDIO
UpdatedMay 28, 2026
