Ruby & Rails LLM Discoverability Scorecard

01The scorecard

Measured over HTTP, June 2026, against each project's documentation page (✓ good, ✗ missing). Retrieval: can an agent read the docs at request time? Training: would they reach a model's corpus, through code or through the web? The code column holds the corpus's own answer: every docs repo is checked against the official Am I in The Stack? index. A repo already in The Stack skips the web filters, so its web cells dim. Click any heading to sort; what each column means →

Showing 93 of 93

Resource (docs) ▲	Retrieval can an agent find it at request time						Training would it reach the corpus
Resource (docs) ▲	robots allows AI ▲	crawlable (no WAF) ▲	sitemap ▲	llms.txt ▲	content negotiation ▲	.md routes ▲	in The Stack v3 observed; docs repo license ▲	Common Crawl ▲	quality best of 5, ≥3 ▲
Core (Ruby Central / Rails Foundation / community-run)
Ruby (language)	✓	✓	✓	✓	✗	✗	✗NOASSERTION	4,756/—	✓3.7
Rails Guides	✓	✓	✗	✗	✗	✗	✓MIT	276/—	✓3.8
Rails API	✓	✓	✗	✗	✗	✗	✓MIT	1,845/—	✓3.6
RubyGems Guides	✓	✓	✓	✓	✗	✗	✓NOASSERTION = CC-BY-SA (Creative Commons Attribution-ShareAlike)	12/98	✓2.8
Bundler	✓	✓	✓	✗	✗	✗	✓(no license file)	212/—	—
RubyDoc.info	✗blocks ccbot, gptbot, claudebot, google-extended, applebot-extended	✓	✗	✗	✗	✗	—	4/—	✓3.8
Frontend & View
Hotwire	✓	✓	✓	✗	✗	✗	✓(no license file)	6/—	✗1.9
Turbo	✓	✓	✗	✗	✗	✗	✓(no license file)	5/—	✗2.0
Stimulus	✓	✓	✗	✗	✗	✗	✓(no license file)	3/—	✓2.8
ViewComponent	✓	✓	✗	✗	✗	✗	✓MIT	12/—	✓2.8
Phlex	✓	✓	✓	✗	✗	✗	✓MIT	44/—	✓3.1
Ruby UI	✓	✓	✓	✓	✗	✗	✓(no license file)	96/—	—
Inertia Rails	✓	✓	✓	✓	✗	✓	✓MIT	21/60	✗1.9
Lookbook	✓	✓	✗	✗	✗	✗	✓MIT	16/—	✗1.6
Vite Ruby	✓	✓	✗	✓	✗	✗	✓MIT	6/—	✗2.1
Web Frameworks
Roda	✓	✓	✗	✗	✗	✗	✓MIT	6/—	✗2.0
Sinatra	✓	✓	✗	✗	✗	✗	✓(no license file)	37/—	✗2.2
Hanami	✓	✓	✓	✓	✗	✗	✓(no license file)	14/—	✗1.9
Bridgetown	✓	✓	✓	✗	✗	✗	✗opted out	68/139	✗2.1
Jekyll	✓	✓	✓	✗	✗	✗	✓MIT	40/210	✗1.9
Rage	✓	✓	✓	✓	✗	✗	✗after cutoff	6/35	✗1.7
Data & ORM
Sequel	✓	✓	✗	✗	✗	✗	✓MIT	125/—	✗2.0
ROM	✓	✓	✓	✓	✗	✓	✓(no license file)	15/—	✗1.7
dry-rb	✓	✓	✓	✓	✗	✗	✓(no license file)	36/—	✗1.8
AI
RubyLLM	✓	✓	✓	✓	✗	✗	✓MIT	65/—	✗2.4
Background, Realtime & Deploy
Sidekiq	✓	✓	✗	✓	✗	✗	✓LGPL-3.0	3/—	✗1.9
AnyCable	✓	✓	✓	✓	✗	✗	✓MIT	36/85	✗2.0
Kamal	✓	✓	✗	✗	✗	✗	✓(no license file)	32/—	✗1.5
Falcon	✓	404	✗	✓	✗	✗	✓MIT	3/—	✗2.0
Karafka	✓	✓	✓	✓	✗	✓	✓NOASSERTION = custom, 'All Rights Reserved' for original wiki content	136/209	✓2.6
Heroku	✓	✓	✓	✓	✓	✓	—	0/—	✗1.7
Fly.io	✓	✓	✓	✓	✗	✗	✓Apache-2.0	14/—	✗1.7
Render	✓	✓	✓	✓	✓	✓	—	0/—	✗2.1
Railway	✓	✓	✓	✓	✓	✓	✓MIT	0/—	✗2.2
DigitalOcean	✓	✓	✓	✓	✓	✗	—	0/—	✗1.8
AWS Elastic Beanstalk	✓	✓	✓	✓	✓	✓	✓CC-BY-SA-4.0	46/—	✗2.4
PlanetScale	✓	✓	✓	✓	✓	✓	✓Apache-2.0	0/—	✗2.3
Supabase	✓	✓	✓	✓	✓	✓	✓Apache-2.0	0/—	✗2.1
Neon	✓	✓	✓	✓	✓	✓	✓(no license file)	0/—	✗2.2
Tooling & Types
Sorbet	✓	✓	✓	✗	✗	✗	✓Apache-2.0	21/86	✓2.8
RuboCop	✓	✓	✓	✗	✗	✗	✓MIT	415/1,681	✗1.9
RSpec	✗blocks gptbot	✓	✗	✗	✗	✗	✗opted out	159/—	✗1.6
TestProf	✓	✓	✗	✓	✗	✗	✓MIT	11/—	✗2.1
Pry	✓	✓	✗	✓	✗	✗	✓(no license file)	0/—	✗1.7
Brakeman	✓	000	✗	✗	✗	✗	✓(no license file)	0/—	—
Standard	✓	✓	✗	✓	✗	✗	✓MIT	0/—	✗1.8
Sentry	✓	✓	✓	✓	✓	✓	✓FSL-1.1-Apache-2.0	45/—	✓2.8
AppSignal	✓	✓	✓	✓	✓	✓	—	126/—	✗2.0
New Relic	✓	✓	✓	✓	✗	✗	✓CC-BY-NC-SA-4.0	28/—	✗1.8
Datadog	✓	✓	✓	✓	✓	✓	✓BSD-3-Clause	0/—	✗1.2
Honeybadger	✓	✓	✓	✓	✗	✓	—	27/—	✗2.5
Rollbar	✓	✓	✓	✓	✓	✓	✓(no license file)	0/—	✗1.6
Bugsnag	✓	✓	✓	✗	✗	✗	—	9/—	✗1.6
Scout APM	✓	✓	✓	✓	✗	✓	✓(no license file)	5/—	✗1.4
Skylight	✓	✓	✗	✗	✗	✓	✓CC-BY-NC-SA-4.0	23/—	✗1.5
Better Stack	✓	✓	✓	✗	✗	✓	—	0/—	✗1.9
Papertrail	✓	✓	✓	✗	✗	✗	—	4/—	—
GitHub Actions	✓	✓	✗	✓	✓	✓	✓CC-BY-4.0	0/—	✓2.7
Libraries
GraphQL Ruby	✓	✓	✗	✗	✗	✗	✓MIT	1,804/—	✓2.5
Rodauth	✓	✓	✗	✗	✗	✗	✓MIT	46/—	✗2.0
Action Policy	✓	✓	✗	✓	✗	✗	✓MIT	17/—	✓2.6
Shrine	✓	✓	✓	✗	✗	✗	✓MIT	6/127	✗2.4
Avo	✓	✓	✗	✓	✗	✗	✓(no license file)	168/—	✗1.6
ActiveAdmin	✓	✓	✗	✗	✗	✗	✓MIT	8/—	✗1.8
Ransack	✓	404	✗	✓	✗	✗	✓MIT	2/—	✗1.4
Pagy	✓	404	✗	✓	✗	✗	✓MIT	5/—	✗1.5
Nokogiri	✓	✓	✓	✗	✗	✗	✓MIT	11/22	✗1.8
Faraday	✓	404	✗	✓	✗	✗	✓MIT	0/—	—
Capistrano	✓	✓	✗	✗	✗	✗	✓MIT	6/—	✗2.2
Trailblazer	✓	✓	✗	✗	✗	✗	✓(no license file)	19/—	—
imgproxy	✓	✓	✓	✓	✓	✗	✓(no license file)	141/529	✗1.4
Flipper	✓	✓	✓	✓	✗	✓	✓MIT	43/70	✗1.7
Devise	✓	✓	✗	✓	✗	✗	✓MIT	0/—	✗1.8
Pundit	✓	✓	✗	✓	✗	✗	✓MIT	0/—	✗1.4
CanCanCan	✓	✓	✗	✓	✗	✗	✓MIT	0/—	✗1.4
Community & Resources
GoRails	✓	✓	✗	✗	✗	✗	—	1,167/—	✓2.7
Drifting Ruby	✗blocks gptbot	✗WAF block	✗	✗	✗	✗	—	7/—	✗1.6
RubyEvents	✗blocks ccbot, gptbot, claudebot, google-extended, applebot-extended	✗WAF block	✓	✗	✗	✗	✓(no license file)	2/16,618	✗1.1
Rails at Scale (Shopify)	✓	✓	✓	✗	✗	✗	—	43/116	✓3.3
SF Ruby	✓	✓	✓	✓	✓	✗	—	14/—	✗1.2
Ruby Weekly	✓	✓	✓	✓	✗	✗	—	219/625	✗1.2
Short Ruby	✓	✓	✓	✓	✗	✗	—	279/293	✗1.2
This Week in Rails	✓	✓	✓	✓	✗	✓	—	32/—	✗1.4
Evil Martians	✓	✓	✓	✓	✓	✓	—	114/—	✗1.7
Hotwire Weekly	✓	✓	✓	✗	✗	✗	—	0/—	✗1.2
Write Software Well	✗blocks ccbot, gptbot, claudebot, google-extended, applebot-extended	✓	✗	✗	✗	✗	—	2/—	✓2.8
Thoughtbot	✓	✓	✓	✗	✗	✗	—	43/—	✗1.7
AppSignal Blog	✓	✓	✓	✗	✗	✗	—	333/1,075	✗2.2
Riding Rails	✓	✓	✓	✗	✗	✗	✓(no license file)	57/—	✗1.6
Joe Masilotti	✗blocks ccbot, gptbot, claudebot, google-extended, applebot-extended	✓	✓	✗	✗	✗	✓(no license file)	2/134	✗1.6
Code with Jason	✓	✓	✓	✗	✗	✗	—	165/532	✓2.6
Maintainable	✓	✓	✓	✓	✗	✓	—	330/482	✗1.6
Ruby News	✓	✓	✓	✓	✗	✗	✓(no license file)	203/—	✓3.0

The second gate: a quality filter

Being crawled is the first gate. A classifier is the second. We scored up to five pages per resource with the open FineWeb-Edu filter, which keeps score ≥ 3. Counting each resource's best page, 19 of 93 clear the bar (top: RubyDoc.info 3.8, Rails Guides 3.8, Ruby (language) 3.7); for 49 even the best of five scores below 2. The filter rewards tutorial prose and penalizes reference docs and code, exactly what developers need most. Docs in a public repo skip this filter entirely; that is what the "in The Stack v3" column measures. Why the code channel decides this →

Licenses, as GitHub reports them

The cell prints the license string GitHub's API returns, verbatim, because that string is what tooling reads. It is lossy: every license GitHub cannot match becomes NOASSERTION ("Other"), so a bespoke permissive license and a source-available one look identical there. Across the 70 resources with a public docs repo:

MIT 32
Apache-2.0 4
NOASSERTION 3 any custom or source-available terms
CC-BY-NC-SA-4.0 2
BSD-3-Clause 1
CC-BY-4.0 1
CC-BY-SA-4.0 1
FSL-1.1-Apache-2.0 1
LGPL-3.0 1
(no license file) 24

The reading predicts; the train set decides. The Stack v3 filters file by file, and only the result is observable: 66 of 70 docs repos are in, including karafka/wiki, whose license file says "All Rights Reserved", while karafka/karafka (LGPL) is out and sidekiq/sidekiq (also LGPL) is in. A missing license is no barrier: The Stack keeps unlicensed files. Opting out works: the rspec and bridgetownrb orgs asked to be removed and are absent org-wide. Where we read a license file ourselves, that row's tooltip carries the terms. The full license guide →

02Rails products, tested against the harvest filters

The table above tracks documentation. Models also learn Rails from complete applications, and agentic RL pipelines mine them mechanically. The SWE-bench family selects top-starred repos, derives the test command from CI config, rebuilds the environment in a container, and turns each merged PR that closes an issue and touches tests into one verifiable training task; SWE-bench Multilingual discards about 30% of candidate repos whose tests will not build or run too slowly. Below: the top-starred open-source Rails products, measured against exactly those filters over the GitHub API.

Product	stars	Code corpus would the code reach a pretraining set			RL harvest could a SWE-bench-style pipeline mine it
Product	stars	license as GitHub reports it	SWH archive	in The Stack v3 observed	test CI	median run	container recipe	issue-linked PRs w/ tests of last 50 merged
Maybepersonal finance app; archived 2025-07-24	54,340	AGPL-3.0	✓	✗	✗no test workflow	—	✓	0/50
Mastodonfederated social network	50,182	AGPL-3.0	✓	✗	✓Ruby Testing	14 min	✓	0/50
Huginnself-hosted automation agents	49,766	MIT	✓	✓	✓CI	8 min	✗	0/50
Discoursecommunity forum	47,624	GPL-2.0	✓	✗	✓Tests	14 min	✓	0/50
Metasploitpenetration-testing framework	38,773	NOASSERTION	✓	✓	✓Extended Tests	<1 min	✓	5/50
Chatwootcustomer support platform	35,659	NOASSERTION	✓	✓	✓Frontend Tests	7 min	✓	4/50
GitLabDevOps platform (GitHub mirror; development on gitlab.com)	24,529	NOASSERTION	✓	✓	✗no test workflow	—	✓	0/50
Foremcommunity platform behind dev.to	22,762	AGPL-3.0	✓	✗	✓CI	9 min	✓	3/50
DocuSealdocument signing	18,227	AGPL-3.0	✓	✗	✓CI	5 min	✓	0/3
Postalmail delivery platform	16,725	MIT	✓	✓	✓CI	3 min	✓	2/50
OpenProjectproject management	15,811	GPL-3.0	✓	✗	✓Test suite	24 min	✓	0/50
Diasporafederated social network	13,657	AGPL-3.0	✓	✗	✓CI	19 min	✗	3/50
Fizzykanban board by 37signals	8,044	NOASSERTION = O'Saasy License Agreement	✗	✗after cutoff	✓CI (Main)	3 min	✓	0/50
Canvas LMSlearning management system	6,764	AGPL-3.0	✓	✗	✗no test workflow	—	✓	0/50
Redmineproject management (SVN mirror; patches via redmine.org)	6,014	NOASSERTION	✓	✓	✓Tests	26 min	✗	—
Zammadhelpdesk and ticketing	5,837	AGPL-3.0	✓	✗	✓CI	93 min	✓	2/50
Lagousage-based billing (the Rails API behind getlago/lago)	431	AGPL-3.0	✓	✗	✓Run Spec	9 min	✓	0/50

Copyleft keeps the Rails flagships out of the permissive code corpus. 6 of 17 products are observed in The Stack v3 train set, and every AGPL- and GPL-licensed product is absent, because v3 keeps permissively licensed and unlicensed files. All six that are in report MIT or NOASSERTION terms. Most products also fail the PR-to-issue linkage filter. 6 of 17 merged at least one issue-closing, test-touching PR in the sampled window, and 5 of 17 have GitHub issues disabled (Discourse and OpenProject triage on their own forums; GitLab and Redmine are mirrors of development that happens elsewhere), so a pipeline mining GitHub finds no harvestable instances there. The runtime filter bites too: Zammad's suite runs a 93-minute median, the kind SWE-bench Multilingual discards. The direct workaround is Multi-SWE-RL: its task pipeline is open, it accepts community contributions, and Ruby is absent from it today.

03What will move the needle

Four levers, ordered by depth, all pushing the same number. Rails is plural by design, so the work is shared defaults and shared conventions. Each layer shows its goal as a live gauge; all of them feed the final boss below.

Layer 0: get into the corpus at all ship now

Crawlable, unblocked87/93

Sitemaps56/93

Unblock AI crawlers (CCBot, GPTBot, ClaudeBot, Google-Extended) in robots.txt and at the WAF. One line at RubyEvents frees ~15,775 pages of talks.
Add sitemaps, server-render, link internally, earn backlinks.
CC-license conference video and publish transcripts: Gemini trains on YouTube, and CC-licensed talks flow into open corpora.

Layer 1: win retrieval and publish comparisons (content) ship now

Win retrieval

Why: an agent fetching a doc should get Markdown it can read. Today it gets HTML to scrape.

Content negotiation16/93

.md routes23/93

llms.txt51/93

Serve Markdown by content negotiation and .md routes (Mime::Type.register "text/markdown", :md). Plain HTTP, already used by agents, the durable bet. Ship llms.txt too; it is cheap.
Teach rdoc to emit Markdown and content negotiation by default: one change that lifts every gem at once.

Publish the missing comparisons

A model argues for what its corpus argues for, and almost nothing argues, with numbers, that Rails is the better build for these products. 0 of 12 comparisons are solid; the rest are generic takes or missing.

Build … in Rails	vs JS fullstack Next.js / Node	vs Python FastAPI / Django
B2B SaaS multi-tenant, team invites, billing	missing	missing
Online store catalog, cart, checkout, admin	generic¹	generic¹
Team inbox shared inbox, collaborative replies	missing	missing
Issue tracker projects, issues, statuses, comments	generic²	missing
Approval workflow upload, route for sign-off, track	missing	missing
Internal admin staff-login CRUD dashboard	generic²	generic³

Sources: ¹ monterail.com · ² kunalganglani.com · ³ goudeketting.nl

Snapshot 2026-06: one web search per task and stack, then judged. solid current, task-specific, with numbers · generic framework pros and cons · missing nothing credible.

Layer 2: make agents fluent in the gems (tools) ship now

Shared gem/agent conventionnone yetAgent Skills conventionfragmented

Agree on a convention so a gem ships agent tooling (an MCP endpoint, a skill) the way it ships a README.
Converge on one Rails MCP server: agents introspect the app (gems, versions, schema, routes) and pull current docs on demand.
Agree on a shared Agent Skills convention so skill packs interoperate.
Copy Laravel Boost: official MCP, version-pinned guidelines, on-demand skills. Rails has the parts (fast-mcp, Tidewave, rails-mcp-server, the official Ruby MCP SDK), and a Boost-shaped bundle is emerging in rails-hyperdrive new.

Why: models know standard Rails. The gems, and anything past the training cutoff, are where agents guess. A convention scales the fix across that long tail.

Layer 3: change the training default long game

Ruby in Multi-SWE-benchabsentOpen idiomatic-Rails datasetnone yet

Contribute real Rails repos to Multi-SWE-bench and publish an open idiomatic-Rails eval. Agentic benchmarks are where coding ability is now measured, Ruby is absent from them, and adding a language to an eval measurably lifts models on it (MultiPL-T, Bridge-Coder).
Grow Ruby's share of the corpus. In The Stack v2, Ruby is ~6.8 GB to Python's ~60, and capability tracks share. Publish an open idiomatic-Rails instruction dataset; contribute permissive Ruby to open corpora like Common Corpus.
Re-run the public whichlang benchmark on each new model; it is the scoreboard below.

★ The final boss

Frontier models reach for Ruby more often. One metric, fed by every layer above: given a free choice, 13 models picked Ruby 0 times in 1,267 solutions (they default to Python, JavaScript, and Go). Win condition: the zero starts climbing, model after model.

Ruby picks0/1,267

04Methodology

Probed over HTTP, June 2026, against each project's documentation URL: robots.txt parsed for AI user-agents; crawlability fetched as CCBot, which catches WAF blocks; content negotiation asked with Accept: text/markdown; .md routes and llms.txt checked for a 200. The Stack v3 membership is read from the official Am I in The Stack? index, one lookup per repo owner. An absence is attributed only when the reason is verified (an opt-out issue, a repo public after the 2025-08-07 crawl cutoff) and says "not established" otherwise. The language-choice figure is whichlang's: 13 models, 1,267 classified solutions, 0 Ruby. The Rails-products table is measured in August 2026 over the GitHub API (stars, license, workflows and their run durations, the last 50 merged PRs via GraphQL), the Software Heritage origin API, and the same Am-I-in-The-Stack index.

Why Common Crawl? It seeds most open web corpora (C4, FineWeb, behind GPT and Llama), samples by domain centrality, and never copies a site in full, so coverage is a proxy for whether a model saw the docs at all. It also reflects crawls already taken: the one column you cannot fix this quarter, which is why Layer 0 exists. How it samples →