Methodology

How the Agent Index is scored — formula version v1. The weights and constants below are versioned in our git repository; any change bumps the version and is noted here.

The integrity policy

No pay-for-placement, ever. There is no “featured”, “sponsored”, or “promoted” field anywhere in the Index data model — a paid position is not just against policy, it is structurally impossible.
Sponsors, when they exist, buy labeled media — never rank. Money is not an input to any formula on this page.
Scores are recomputed live from raw data on every view. Nothing is hand-adjusted; changing the formula changes it for every agent at once, in public.
CodexMaster is independent and not affiliated with any tool it ranks.

The formula (v1)

Each agent's score is a weighted blend of three components, each normalized to 0–1, totaled to 100:

score = 100 × ( 0.35 × pricing + 0.3 × cadence + 0.35 × vibe )

Component	Weight	Normalization
Pricing (value)	35%	1 − (typical operator spend / $200/mo), floored at 0. Free and open-weight agents score 1. "Typical spend" is what a serious daily operator pays, maintained by hand from public pricing pages.
Release cadence	30%	Releases logged in the last 90 days ÷ 12, capped at 1 (≈ weekly shipping earns full marks). Release entries are manual at P0; agents with no entries yet score a neutral 0.5, marked "provisional".
Community vibe	35%	Mean of the last 30 days of better/same/worse votes (+1/0/−1) mapped to 0–1, shrunk toward neutral 0.5 by a 10-vote prior so a handful of votes can't swing the Index. No votes yet ⇒ neutral 0.5.

The v1 weights are launch defaults: the two hard-data components (pricing + cadence) deliberately outweigh community vibe while vote volume is young. They will be revisited — and the version bumped — as the vote stream matures. Every agent page shows its exact component arithmetic, so you can always answer “why is this #3?”.

How votes are handled

One click: is this agent better, the same, or worse than yesterday? Votes count +1 / 0 / −1.
Anonymous is fine. Visitors are identified only by a salted hash of IP + browser signature; the raw values are never stored.
One vote per agent per visitor per day, enforced by a database constraint — not just the UI.
Votes are an immutable, append-only event log. Rows are never edited or deleted (the database physically refuses), so any score can be recomputed from raw history at any time.
Anti-abuse is automated, not moderated: 8+ votes from one visitor within 60s, or more than 20 votes in a day, are flagged at insert time and excluded from every aggregate (they remain in the log). Beyond 40/day, votes are rejected outright.

Where the data comes from

Pricing — maintained by hand from public pricing pages, with an append-only change log per agent. Entries still awaiting verification are labeled “data unverified” on the Index.
Releases — entered manually at this stage (automated feeds later). An empty release radar scores neutral, never zero, so no agent is punished for our data lag.
Vibe — you. The vote stream is the only component vendors can't touch and can't buy.

Spot an error? Corrections beat complaints: tell us and we'll fix the data, publicly.

Methodology

How the Agent Index is scored — formula version v1. The weights and constants below are versioned in our git repository; any change bumps the version and is noted here.

The integrity policy

No pay-for-placement, ever. There is no “featured”, “sponsored”, or “promoted” field anywhere in the Index data model — a paid position is not just against policy, it is structurally impossible.
Sponsors, when they exist, buy labeled media — never rank. Money is not an input to any formula on this page.
Scores are recomputed live from raw data on every view. Nothing is hand-adjusted; changing the formula changes it for every agent at once, in public.
CodexMaster is independent and not affiliated with any tool it ranks.

The formula (v1)

Each agent's score is a weighted blend of three components, each normalized to 0–1, totaled to 100:

score = 100 × ( 0.35 × pricing + 0.3 × cadence + 0.35 × vibe )

Component	Weight	Normalization
Pricing (value)	35%	1 − (typical operator spend / $200/mo), floored at 0. Free and open-weight agents score 1. "Typical spend" is what a serious daily operator pays, maintained by hand from public pricing pages.
Release cadence	30%	Releases logged in the last 90 days ÷ 12, capped at 1 (≈ weekly shipping earns full marks). Release entries are manual at P0; agents with no entries yet score a neutral 0.5, marked "provisional".
Community vibe	35%	Mean of the last 30 days of better/same/worse votes (+1/0/−1) mapped to 0–1, shrunk toward neutral 0.5 by a 10-vote prior so a handful of votes can't swing the Index. No votes yet ⇒ neutral 0.5.

How votes are handled

One click: is this agent better, the same, or worse than yesterday? Votes count +1 / 0 / −1.

Anonymous is fine. Visitors are identified only by a salted hash of IP + browser signature; the raw values are never stored.

One vote per agent per visitor per day, enforced by a database constraint — not just the UI.

Votes are an immutable, append-only event log. Rows are never edited or deleted (the database physically refuses), so any score can be recomputed from raw history at any time.

Anti-abuse is automated, not moderated: 8+ votes from one visitor within 60s, or more than 20 votes in a day, are flagged at insert time and excluded from every aggregate (they remain in the log). Beyond 40/day, votes are rejected outright.

Where the data comes from

Pricing — maintained by hand from public pricing pages, with an append-only change log per agent. Entries still awaiting verification are labeled “data unverified” on the Index.

Releases — entered manually at this stage (automated feeds later). An empty release radar scores neutral, never zero, so no agent is punished for our data lag.

Vibe — you. The vote stream is the only component vendors can't touch and can't buy.

Spot an error? Corrections beat complaints: tell us and we'll fix the data, publicly.