Features

What Hover does today, across its four surfaces — authored through the MCP, reviewed in the optional cockpit, run in CI.

Authoring — the MCP

The authoring engine is an MCP server (@hover-dev/mcp) you add to the coding agent you already run. Once it's wired in, your agent gets Hover's grounded browser tools plus a test_app prompt that explores your app and crystallizes a suite.

• The MCP tools — the grounded actuation tools (click_control / fill_control / select_control / check_control), browser_navigate / browser_snapshot, crystallize_spec, the recall / record_fact memory tools, and the test_app / heal prompts. The agent acts through these so the selector that drove an action is the exact one saved.
• API testing — Hover passively captures the app's browser traffic off the same CDP connection (no proxy) and crystallizes the checks worth pinning into a plain *.api-test.spec.ts — verified with replay_request so no status code is confabulated.
• Login & credentials — the agent logs in via a per-environment @account; the password is redacted to process.env.HOVER_PASSWORD, and the login is lifted into an auth.setup.ts fixture that runs once for the whole suite. Never written into the spec, the JSDoc, or the sidecar.
• white-box source reading — an opt-in, read-only source reader the agent can use to ground its testing in your code. Off by default; fenced to read-only.

Output

• Crystallize a spec — __vibe_tests__/<slug>.spec.ts using getByRole / getByLabel / getByTestId, emitted deterministically (no LLM writing code). Runs in CI without Hover. Every interaction is wrapped in a visibility prelude so UI drift fails in ~3 s with Locator expected to be visible instead of a 30 s actionability timeout.
• Structured spec output — the saved spec grows an architecture: test.step stages, popup / new-tab Promise.all pairing, and a .hover/sidecars/<slug>.json sidecar.
• Page Object extraction — when a NON-login flow recurs across specs, Hover lifts it into pages/ + fixtures.ts and folds the specs, so a change to that flow is a one-file fix. Deterministic, no LLM.
• Optimize a spec — an optional, off-by-default AI pass polishes a deterministic draft into a diff-reviewed candidate: adds the assertions the run observed, flags buggy behaviour // KNOWN BUG, always keeps the original.

Maintenance

• Self-heal a drifted spec — when the UI changes and a recorded selector stops matching, /mcp__hover__heal replays the spec against the live app, finds the broken step, and re-grounds only that step. Record == replay is preserved.

Memory

• Test knowledge that compounds — Hover keeps a Business Map of your app's flows + remembered rules under .hover/, committed with your code. The suite stays self-aware and gets smarter as your app grows.

Review — the optional cockpit

The VS Code extension (hyperyond.hover-dev) is an optional review cockpit, not the engine: a Business Map graph of your flows + coverage, and a Dashboard (pass / fail / flaky + CI results), with one-click run. It reviews — it drives no agent.

Run — CI

The crystallized specs are plain @playwright/test. They run on every PR with no agent, no tokens, no key. Hover can generate the GitHub Actions workflow and pull the run results back into the Dashboard.