What AI does
AI analyzes tasks, changes code, adds tests, runs checks, and investigates failures. The work continues across modules.
AI-autonomous software engineering experiment
ZeroWeb tests whether AI can sustain a large systems project. One human designs, decides and verifies; AI develops the Rust browser and its product code, with almost no code written by hand.
Experimental · Not ready for everyday browsing or production use
The experiment
A browser is large enough to test more than code generation. It requires long-term architecture work, standards implementation, debugging, security boundaries, graphics, and cross-platform integration. Verification runs throughout.
AI analyzes tasks, changes code, adds tests, runs checks, and investigates failures. The work continues across modules.
Specifications, module boundaries, review rules, CI, WPT, reftests, and browser comparisons decide whether a change is acceptable.
A larger repository is not the goal. Features must compile, run, and pass the checks appropriate to the behavior being changed.
The result is judged by working software and reproducible checks, not generated line count.
Runnable code
The desktop entrypoint used to integrate and test the browser subsystems. It is not yet an everyday browser.
View browser applicationA Rust API for embedding the page pipeline in host applications. The API and compatibility surface are still evolving.
View WebView libraryImplementation scope
The repository separates the systems that load, parse, lay out, and render a page so each part can be implemented and verified independently.
Engineering method
The browser core, host layers, and product code are written in Rust. External crates provide bounded low-level capabilities.
Web behavior is verified against WPT, CSSWG reftests, and browser comparisons.
Formatting, static analysis, unit tests, integration tests, and platform CI guard the repository.
Updates link to commits and roadmap items so the experiment can be inspected rather than taken on trust.
Current status
Core subsystems and runnable entrypoints exist. Real-site compatibility, complete DOM and Web APIs, and everyday browser behavior remain active work.
See the detailed roadmapRun the code
The repository includes a desktop browser entrypoint and a smaller WebView rendering demo. Platform prerequisites are documented in the README.
Launch the desktop browser entrypoint.
$ make browserRun the smallest rendering pipeline demo.
$ cargo run --bin webview-demoThese commands build experimental software; see the README for platform prerequisites and V8 setup.
Read the setup guide →Project journal
A running record of what changed, what is still in motion, and where each update comes from.
Seven of the twelve closed goals were chartered during the same week, and webcrypto-compat-v3 went from charter to archive in two days. The full-WPT pass rate rose from 42.11% to 68.36%. Rendering failures fell from 757 to 691.
Read update →Six goals closed with final numbers, including css-testharness-compat, chartered on 10-27 and closed on 10-30. As the large encoding and wasm corpus runs were recorded, the full-WPT pass rate rose from 7.41% to 42.11%. Rendering failures fell from 884 to 757.
Read update →33 previously closed goals were archived, and 15 more reached a terminal state. Rendering fixes reduced full-corpus failures from 1,104 to 884; media timing and browser stability also improved.
Read update →Inspect the experiment
The repository, development history, tests, roadmap, and weekly updates are public. Contributions are expected to meet the same verification rules.
A browser gives AI development a long-running test case: architecture, code, debugging and verification must work together.
Follow the page pipeline and the boundaries between the browser application, embeddable WebView, engine and rendering infrastructure.
Prepare a local checkout, launch the desktop browser and inspect the WebView API before embedding the experimental engine.
Understand self-source reftests, assertion-based suites, changing corpus sizes and p95 benchmarks before comparing results.