Methodology

Five-star ratings are broken — they hide how they were formed. The ToolScore Index is 10 criteria × 10 points, each score visible, each number sourced and dated. If a tool scores 61/100, you can see exactly which pillars cost it points. The person and the receipts behind these numbers: about.

💰 The real bill, not the marketing one

Hours of work per dollar, real limits, the cost of the whole stack.

🧠 The “why”, not just the ranking

Every pick carries the reason an insider would give you over coffee.

⚠️ No hidden traps

The fine print, the promo price that jumps, the “free” that isn't.

Who's behind these numbers →

The 10 criteria

Criterion	What 10/10 looks like	What 0–3 looks like
Price / Value	Free or cheap with real capability; predictable cost	$40+ tiers, metered surprises, teaser entry price
Token & Rate Limits	A plan holds a full working day of heavy use	$20 dies in 2–3 hours of active work
Speed / Latency	Streaming, responsive, interactive	Multi-second stalls, queueing
Privacy & Retention	No training on your code by default; zero-retention or self-host	Training by default, opt-out buried
BYOK	Bring your own API key; pay raw token prices	Locked to vendor's model menu and markup
Output Quality	Top-tier benchmarks and real-world task closure	Plausible-but-wrong code, weak reasoning
Stability	Predictable behavior, uptime, few regressions	Broken sessions, breaking changes
Ease of Setup	Minutes from install to first real task	Config archaeology before productivity
Billing Transparency	You can predict the invoice before it arrives	Fuzzy windows, credit meters, moving caps
Support & Community	Docs, responsive channels, active ecosystem	Community-only, stale docs

Verdicts: 80+ Best Value · 65+ Strong · 50+ Average · <50 Weak.

Where the data comes from

Our collection pipeline — we call it ToolRadar — assembles every topic from two kinds of sources, always labeled on the page:

Official vendor pages — pricing pages, docs, changelogs; curated per topic and fetched directly from the vendor. Dated, because they change.

Community evidence — Hacker News threads, GitHub issues, Reddit when reachable: where the real pain points live. Never presented as our own testing. Each fact carries its date, and GitHub evidence also carries the issue state (open/closed) and who said it — maintainer, contributor, or user.

Off-topic material is stopped at the door: a thread that never mentions the topic doesn't enter the corpus, so Sources sections stay clean. We don't run lab benchmarks and we don't pretend to — when something is a single user's report, the article says so.

Fact verification

Nothing prints on faith. Two mechanisms, depending on what you're reading:

Tool scorecards — each fact is re-searched against independent sources and published with its source link; facts that can't be confirmed are labeled unconfirmed or cut.

Deep-dive articles — every fact must arrive with a verbatim quote from a collected source: no quote, no fact. In the finished text, every number must trace back to that collected evidence — an automated check that forces a rewrite when it doesn't (twice, then a human). And a comparison isn't written at all unless at least two-thirds of the tools named in the headline actually have data behind them.

Attribution instead of aggregates: you'll read “maintainer confirmed (Aug 2026)” or “a single user report, 2024” — never invented composite scores. Community evidence older than a year is explicitly dated (“as of 2024”), and facts from closed or resolved issues are never described as current problems.

A site that debunks hollow data can't publish unverified numbers as facts — that constraint shapes everything here. Nothing publishes automatically: a person reads every draft before it goes live. If you find a stale number, report it; we'd rather correct than defend.

Scores vs. affiliate links

ℹ️ The firewall

Affiliate relationships never touch scores. Tools without any affiliate program regularly top our rankings — that's the proof. Money details live in the disclosure; the scoring pipeline has no knowledge of them.

Freshness

Every number carries its date in the text itself. Official pages are re-fetched on a rolling refresh window; community evidence keeps the date it was said, and anything older than ~12 months is dated in prose (“as of 2024”) so you can weigh it yourself. Tool scorecards carry a “data as of” date and are re-checked when their facts move. If something looks stale, treat it as suspect — including by us — and say so.