ReachOS is a Chrome extension that scores your tweets in real time against 25 signals (the 22-signal taxonomy xAI published on 2026-05-15, plus topic consistency, ad disclosure, and post frequency) and shows you what to change. A numeric reach interval is shown only after your own posts beat your usual reach on held-out history. Until then the forecast says uncalibrated and the checklist is the output. BYOK (Bring Your Own Keys) — fully self-hostable.
You type a tweet.
ReachOS scores the checklist as you type.
Until your history beats your usual reach, the forecast says "uncalibrated".
v4.0 (May 2026): ReachOS realigned to xAI's published Phoenix-era algorithm. Outbound links are no longer penalized; instead, descriptive link anchors with a curiosity gap drive the new click signal. See CHANGELOG.md for the full migration notes.
Features
Score Overlay + Reach Forecast
Real-time 0-100 score as you type in the X.com composer. Signal bars grouped into 4 buckets — Engagement, Curiosity, Dwell, Risk — summing 25 signal predictors. Reach Forecast shows an interval only when a per-account model beats that account's rolling median on newer posts. Otherwise it stays uncalibrated.
X-Ray Mode
Scores every tweet on your timeline as you scroll. Color-coded pills (red/orange/yellow/green/blue/purple) show reach potential at a glance.
AI-Powered Analysis (BYOK)
- AI Slop Detection - Flags AI-sounding language (28 weighted patterns + Claude verification)
- Hook Quality Assessment - 6-dimension analysis of your opening line
- Auto-Optimize - 5 rounds of iterative AI rewriting, keeps the best version
- Self-Reply Generator - Creates a self-reply to kickstart conversation (150x algorithm boost)
Smart Signals
- Trending Alignment - Detects when your tweet matches trending topics (+5 bonus)
- Posting Time Optimizer - Shows whether now is a good time to post based on research-backed UTC windows
- Reply Coach - Notifies you about unanswered replies on tracked tweets
- Account Health - 5-factor health score with actionable tips
Self-Learning Loop
For signed-in users, the system gets smarter over time:
Post tweet → Save prediction → Fetch real metrics (15min)
→ Fit a per-account model on older posts, test it on newer posts
→ Show a reach interval only if that model beats your rolling-median reach
→ Otherwise the forecast stays uncalibrated and the checklist remains
→ Learn which signals lift YOUR tweets (daily, needs 5+ tracked tweets)
→ Your Reach Score applies those learned weights (±6 points per signal)
How It Works
Scoring Engine (25 Signals)
ReachOS v4.2 mirrors the 22 signals scored by home-mixer/scorers/ranking_scorer.rs in xai-org/x-algorithm, plus 3 ReachOS-original predictors covering ad-disclosure, daily post-frequency decay, and microniche topical consistency. Each signal has its own predictor file in packages/rules-engine/src/signals/ — and every xAI signal name matches the Rust source exactly. A reach number is a per-account model judged on newer posts; if it does not beat that account's rolling median, the forecast stays uncalibrated and the checklist is the output.
| Bucket | Signals | What it predicts |
|---|
| Engagement | favorite, reply, retweet, quote, share, share_via_dm, share_via_copy_link, click, photo_expand, vqv, quoted_click, quoted_vqv | Direct positive interactions X scores explicitly |
| Curiosity | profile_click, follow_author | "Who is this person?" → audience-growth signal |
| Dwell | dwell, cont_dwell_time, cont_click_dwell_time | Reader stays vs. scrolls past |
| Risk | not_dwelled, not_interested, block_author, mute_author, report | Negative actions X scores as soft/hard penalties |
Base score 30 + Σ positive signal predictors (max 95) + Σ negative signal predictors (min −50), clamped to 0–100. Five signals are conditional: photo_expand (needs image), vqv (needs video), cont_click_dwell_time (needs clickable content), quoted_click / quoted_vqv (need quote-tweet).
Key signal weights (mirror of ranking_scorer.rs)
| Signal | Max points | Why this weight |
|---|
reply | +12 | X's king positive signal (confirmed ~27× like in 2023 leak; retained in May 2026) |
click | +8 | Replaces v3's link penalty. X scores URL clicks as a positive engagement |
dwell | +8 | Single-view dwell, scored independently of cont_dwell_time |
favorite | +7 | Likes — broad signal but lower per-action than reply |
profile_click | +7 | Curiosity → follow funnel |
retweet | +7 | Quotable, succinct content |
share_via_dm | +6 | NEW in May 2026: DM-share scored as its own signal |
quote | +6 | Quote-tweet-worthiness (contrarian claims drive this) |
share | +5 | Generic broadcast share |
follow_author | +5 | Follow conversion |
photo_expand | +4 | Conditional on image attached |
vqv | +4 | Conditional on video attached |
not_dwelled | −12 | Scroll-past risk |
report | −15 | Hardest negative (mirrors ~−738× like in 2023 leak) |
What's removed from v3.0: outbound-link penalty, 3+ hashtag penalty, unknown-language penalty, competitor-mention penalty. None have any visible representation in X's May-2026 published code.
Reach Forecast
The checklist score is not a reach prediction. A number is a separate model, and it is withheld when that model does not beat the account's recent typical impressions.
The forecast fits a ridge regression per account:
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
