Run an evolving component in your browser

Preview. Open the browser workspace to change a marketing component's content and layout, inspect a suggestion, and save the version you choose. Its history stays in your browser. You can use it without an application server, account, provider key, or cloud inference.

IndexedDB stores application records, Web Locks coordinate writes across tabs, and the Rust expression evaluator runs as WebAssembly. The renderer displays a limited set of component elements. The browser, renderer, storage adapter, and evaluation policy are trusted host software that the component cannot rewrite.

Try a change

Open the workspace.

Choose an audience and release stage.

Select Save context.

Select Suggest a local change to generate a suggestion from local rules.

Inspect the proposed content, fit score, and checks across four contexts.

Select Use this version when the suggestion passes. It becomes visible after the browser adapter saves the new application state.

Reload to reopen the same application.

You can restore earlier behavior as a new history step while keeping the current signals. Export saves a portable copy of the history; import opens a verified copy in a fresh workspace.

Each workspace has its own URL. Create workspace creates an application at its starting version and preserves earlier records. Use Back or a bookmarked workspace URL to return. Start another workspace when an experiment reaches its history or candidate limit.

Automatic mode makes one attempt using local rules after you save context. It adopts the suggestion only if it passes the checks and strictly improves the fit score. The mode starts off in every tab. Pause and pin controls are saved with the application. Inference does not run in the background.

The fixed content and layout policy scores how well a suggestion matches your chosen audience and release stage. This score measures only those criteria; conversion, factual accuracy, general writing quality, and accessibility across arbitrary content are outside its scope. The model cannot change the score or bypass the independent checks across contexts.

Use on-device inference

Download & load local AI loads a pinned model in a dedicated WebGPU worker, which runs inference separately from the page. Model files total about 203 MiB, plus runtime, transfer, and temporary/cache overhead. The button is unavailable when the browser lacks the required GPU features. Ordinary application use does not download the model.

The model can suggest the same content and layout fields as the local rules. The application checks its output before offering it for adoption. A suggestion may parse successfully and still fail the checks or fail to improve the fit score. Stop or timeout terminates the worker. Another attempt requires you to load it again; uncertain attempts are never retried automatically.

See browser inference for model versions, compatibility, file-integrity checks, and results from the tested browser and device. Model caching and shader compilation are separate from saving application history.

Save and reopen offline

IndexedDB stores the application history. It saves content-addressed records (records identified by hash), prepared operations, and the pointer to the current state in separate steps. Web Locks keep one writer in control during asynchronous checks and writes. The adapter requests strict transaction durability, a browser hint that cannot guarantee protection against device failure. Commands based on an older state and conflicting operation identities are rejected before they can overwrite another tab's work.

localStorage remembers only the selected workspace. The /living/ editor uses separate preview storage; its previews are not migrated into application history.

A service worker scoped to /grow/ caches the application page, evaluator, fonts, and scripts. The build pins each file's bytes. An incomplete or mismatched update cannot replace the installed version. Save for offline use checks these files and requests persistent browser storage. The interface reports whether the browser granted that request. Other website routes require a connection.

Offline readiness covers the application and local rules. AI also needs its model and optional worker file cached, GPU support, and sufficient resources; downloading the model alone does not establish readiness. Browser storage may be cleared or evicted, and private sessions may be temporary. Keep an export for data you care about.

History and recovery

The application saves program versions, memory, execution records, candidate evaluations, and the evidence linking each adoption to its starting state. A workspace holds at most 64 states and 16 candidate evaluations. Signals and control changes also consume states. Missing or corrupt dependencies stop the operation; later execution cannot silently replace the missing records.

Before an imported workspace is selected, its export is checked against the saved current state and replayed. Exports establish content integrity and pure replay. They carry no author signature or permission to perform external actions. Import creates a fresh database and preserves the current one. If ordinary opening fails, recovery export saves raw records for diagnosis. Those records have not passed the application-export checks. Storage quota errors appear in the interface and leave previously saved state intact.

Development and browser tests

For local development, run bun run check, then serve site/dist on localhost. A secure context, IndexedDB, and Web Locks are required to change application state. WebLLM is a build dependency; the ALGAL CLI has zero required runtime dependencies. Browser records use the same canonical SHA-256 identities as Bun and the native runtime.

scripts/browser-grow-qualification.mjs tests IndexedDB and Web Locks, interrupted writes, immutable operation identities, corrupt records, competing tabs, import, restoration, and reload/adoption with networking disabled. CI installs its pinned Chromium and runs the local-rule workflow. Set ALGAL_PLAYWRIGHT_MODULE and ALGAL_BROWSER_EXECUTABLE to installed tools; ALGAL_BROWSER_EVIDENCE selects the test profile and results directory. ALGAL_GPU_QUALIFY=1 also downloads the pinned model, generates a suggestion, and reloads the cached model with networking disabled.

The browser inference test record lists the tested device, browser, and offline model results.