Project
Status
What is measured, what is proven in use, and what isn't yet. Updated with every release; the numbers come from the repository's own data.
Version 0.2 (November 2026). Usable for prototypes and small apps built with an AI agent. The syntax may still change before 1.0; every change is listed in the changelog with what to write instead.
Measured #
- Cost: over 50 tasks and three models, ArtScript cost 48% less per working result than React + TypeScript with Claude Sonnet 5.5, 45% less with Opus 5.5 and 36% less with Haiku 4.5, and less than Svelte, Vue and Solid with each. Against Next.js 16, measured on 2026-11-04: 50%, 45% and 29% less (64%, 59% and 37% on the full-stack tasks). Sonnet and Opus solved 100% of the tasks in ArtScript and in Next.js; with Haiku, ArtScript solved 90% and Next.js 96%, so Next.js fails less with the smallest model. Raw data and method: Benchmarks.
- Performance: in the official js-framework-benchmark results (Chrome 154), ArtScript scores 1.32 where 1.00 is the fastest keyed implementation: between Svelte (1.24) and Vue (1.33), ahead of React (1.91); 7.5 KB compressed. Apps ship 5–20 KB of JavaScript.
- Tests: 287 in CI, covering the compiler, the runtime, the server, the dev tools and every example and guide on this site.
Proven in use #
- Three apps built by AI agents with the published package and only the spec: a shop, a backoffice and a landing page. Every bug they found is fixed.
- This website is an ArtScript app.
- The three apps run as static demos (their backend in your browser). No ArtScript server has run in production with real users yet.
Not yet #
- Security audit. The server has CSRF protection, rate limits, security headers, hashed passwords and sessions that expire, and the security policy explains what apps get by default. Nobody outside the project has audited it. Don't put other people's sensitive data in an ArtScript app before that happens.
- Frozen syntax. Planned for 0.3. Until then, a release can change the language;
art fmtand the changelog tell what to write instead. - Small models. With Haiku 4.5, one task in ten needed a human. Use Sonnet or Opus for real work.
- Imperative code. The saving is in apps of UI and data: pages, state, the api, forms, lists, tests. In a physics loop, a parser or canvas drawing, ArtScript is JavaScript with another syntax for statements and saves nothing; put that code in a
.tsfile imported withuseand expect the same cost as with any framework. Data, constants and helpers stay in.art:art checkwarns about a module that could be ArtScript. The benchmarks page has a task of that kind measured apart. - Database: SQLite only, one process. Every field is indexed (measured: inserts and
where/sortqueries stay at 0.3 ms and 2 ms with 64,000 rows;searchscans the table). Enough for small and medium apps; no Postgres yet. - Known gaps: prerendered pages are replaced, not hydrated, when the app loads (a brief flash); server functions aren't transactional; with
api users: User login, any signed-in user can list every account (being decided); no OAuth or email in the static demo mode. - One maintainer (Leandro Bisceglie). Issues are answered, but not within hours.
How to help #
Open an issue with a .art file that shows the problem, or tell us what your agent got wrong and what it wrote: that's how most fixes here started.