Noemium

Pick AI tools without getting played.

AI directories are pay-to-list. Noemium is open source. Every recommendation is a PR you can audit. Receipts where we print a number.

Noemium is a hub for AI practitioners:

• Tools — honest verdicts (ship / situational / skip) with named limitations, evidence links, and a decision briefing (strengths / use_for / skip_when) on flagship cards.
• Agents — operational field guides with install paths, requirements, cost scenarios, security boundaries, and explicit evidence tiers. Grok Bot and Grok Build are treated as separate products.
• Stacks — copy-pasteable recipes: which tools combined, at what monthly cost, for which use case.
• Models — context windows, per-MTok pricing, strengths and anti-use cases, with attributed sources.

Every entry carries last_verified and an observed_by GitHub handle. Tool and stack entries require at least one receipt; agents declare typed evidence tiers (field-tested / source-verified / radar). Affiliate links, when they exist, are always declared.

The site

Beyond the catalog indexes, noemium.com ships:

• /status — build-time catalog health: briefing and receipt coverage, no vanity counters.
• /verified — the field-tested shelf. Empty until earned.
• /refusals — a public ledger of declined paid placements. Zero so far, and that zero is the point.
• /kit — save and share tool sets via URL. No account, no fake totals.
• /changelog — a weekly diff of what changed in the catalog.
• /graveyard — dead tools, with obituaries.
• /why and /method — the essay and the verification method.
• 45 static X vs Y comparison pages, a Pagefind search, and a five-question quiz that ends in a stack.

Open data

• /api/tools.json, /api/stacks.json, /api/models.json — full dumps.
• /tools/<slug>.json — every tool as its own JSON endpoint.
• /llms.txt and /llms-full.txt — the catalog, LLM-ready.

Stack

• Astro (static output) + Preact islands
• Tailwind CSS v4 (CSS-first config, "Bold Grid" design tokens)
• Content: Astro Content Collections (YAML + Markdown), zod-validated
• No WebGL/motion libraries — typographic hero, CSS-only motion

Quick start

npm install
npm run dev        # http://localhost:4321

Other commands:

npm run build      # static build to dist/
npm run check      # astro check (types)
npm run validate   # validate all content entries (CI-friendly)
npm test           # unit tests (schemas, surfaces, kit, logos, harvest)
npm run test:dist  # dist assertions (run after build)
npm run harvest    # observe the world (status/RSS/GitHub/prices/freshness) → reports/harvest/

Harvest

The catalog stays honest only if something watches the world. Harvest is that watcher. It does not patch YAML.

• Registry: src/data/harvest-sources.yaml — what we fetch, how often, how we match it.
• Actions (Mondays 08:00 UTC): deterministic collectors. Statuspage JSON, RSS, GitHub API, LiteLLM/Helicone prices, stale last_verified. Writes an artifact and a single rolling issue, harvest: world watch. No catalog YAML commit. No paid keys.
• World status: npm run world-status writes src/data/world-status.json. Deploy fetches a fresh snapshot before each build (daily 06:00 UTC plus every push) and /signals/ renders it. CI builds the committed file, so tests stay offline.
• Judgement (local): read reports/harvest/latest.md, follow scripts/harvest-judge.md, open a content PR. Merge publishes.

Existing jobs stay: stale-check (adopt-a-page issues), link-check (dead receipts), content-data-refresh (git-derived changelog / price-history). Harvest watches outside the repo.

Content structure

src/content/
├── tools/    # *.yaml — one file per tool (see cursor.yaml as template)
├── agents/   # *.yaml — strict field guides and verdict-free Radar entries
├── stacks/   # *.md   — frontmatter + recipe body (see solo-founder-saas.md)
├── models/   # *.yaml — one file per model (see gpt-5-6-sol.yaml)
└── graveyard/ # *.yaml — dead tools, with obituaries

Schemas live in src/content-schemas.ts and are enforced both at build time and by npm run validate. Agent entries declare field-tested, source-verified, or radar; Radar entries cannot carry verdicts or hard cost scenarios.

Contributing

Every entry is a pull request. See CONTRIBUTING.md — it takes about 5 minutes to submit a tool.

Design decisions are governed by DESIGN.md. Content is licensed CC BY 4.0; code is MIT.