AI visual regression is two phases, not one.

A regular visual regression run pixel-diffs a fresh screenshot against a stored golden PNG, typically via Playwright's toHaveScreenshot(), pixelmatch, or resemble.js. The only knob you tune is maxDiffPixels. AI visual regression replaces that pixel diff with a model that looks at the frame and judges it by meaning: is the Submit button in the right state, did the success toast appear, is the avatar a real image or a gray placeholder. The Assrt codebase has zero calls to toHaveScreenshot, pixelmatch, resemble.js, looks-same, or odiff. The judgment lives in the model, so the test question is English (from scenario.md) and the answer is English (evidence string on each TestAssertion at assrt-mcp/src/core/types.ts lines 13-17).

The default is Claude Haiku 4.5. The exact model ID claude-haiku-4-5-20251001 is declared at assrt-mcp/src/core/agent.ts line 9 as DEFAULT_ANTHROPIC_MODEL. After every visual action (navigate, click, type_text, select_option, scroll, press_key), the browser screenshot is captured and attached to the next tool-result message as { type: 'image', source: { type: 'base64', media_type: 'image/jpeg', data: screenshotData } }. That attach site is at agent.ts line 987. The model sees each new frame, reasons about it against the plan, and decides whether to continue, retry, or emit a pass/fail assertion. No baseline PNG is involved on this path.

After the run finishes, Playwright's native video recorder has written a full WebM of the browser session (stopped at server.ts line 578, finalized and moved into /tmp/assrt/<runId>/videos). You can call the assrt_analyze_video MCP tool with a natural-language prompt. The tool reads the WebM off disk (server.ts line 953), base64-encodes the entire file (line 964), and sends it to Gemini 3.1 Flash Lite Preview as a single inlineData part with mimeType 'video/webm' (lines 976-981). The model replies with an English analysis. You can ask 'did the confirmation banner appear for the full 2 seconds' or 'was there ever a flash of red on the pricing page' and get an answer grounded in the pixels of the run.

Because Gemini accepts video as a first-class modality in a single prompt. The WebM is passed straight in via inlineData with mimeType 'video/webm'. Claude can see images, but sending an entire test recording frame-by-frame burns tokens and misses timing. The assrt_analyze_video tool at server.ts lines 925-1018 is registered only if process.env.GEMINI_API_KEY is set (line 929). If you don't set the key, the tool simply never appears on the MCP server. You still get phase 1 from Claude, just not the retrospective video phase. Nothing else in the codebase changes.

Yes, and that is the intended flow. Call assrt_test to run the scenario (phase 1 happens automatically, Claude Haiku 4.5 judges per frame). The server remembers the path of the WebM in a module-level variable lastVideoFile (server.ts line 270). Now call assrt_analyze_video with any prompt, no videoPath argument needed (server.ts line 939 falls back to lastVideoFile). Repeat with different prompts for free: once the video is on disk, you can interrogate it as many times as you want. 'Did the left sidebar flicker', 'did any modal appear unexpectedly', 'estimate the LCP element visually'. Each question is a different Gemini request against the same base64 blob.

The top results for 'ai visual regression' (Applitools Eyes, BrowserStack Percy, Mabl Visual AI, LambdaTest SmartUI, Reflect) describe AI as a replacement for pixelmatch. They compare rendered screenshots against baselines more cleverly: they ignore dynamic zones, weight structural elements, and suppress false positives. That is one-phase AI. None of them let you take your recorded test run and ask it open-ended questions in English after the fact. The retrospective video analyst is a second surface the industry has not named yet. That is the spine of this page.

The tutorial page at /t/visual-regression-tutorial explains the in-run model in depth: why each screenshot is a JPEG attached to Claude, why there is no baseline PNG, what a TestAssertion looks like. It is a phase-1 deep dive. This page is about the pair. Phase 1 plus the assrt_analyze_video retrospective that the tutorial does not cover. If you care about why baselines are gone, read the tutorial. If you care about the full AI loop (judge during, analyze after), read this.

It stays on disk anyway. The WebM is recorded by Playwright's native page.video() regardless of whether Gemini is configured. It lands at /tmp/assrt/<runId>/videos/recording.webm and a self-contained HTML player is generated next to it at player.html (server.ts line 618). The MCP tool response includes a videoPlayerUrl served by a local static server (server.ts line 629) that auto-opens by default. So even without an API key, you can watch every run replay in any browser. Phase 2 only kicks in if you want programmatic queries over the recording.

Not for everything. Pixel diffs catch sub-pixel layout shifts and one-pixel color drifts a model will not flag. For a design-system component library where a 1px border change is a real regression, keep Playwright's toHaveScreenshot() on the component. Use the two-phase AI approach at the user-journey level: is the checkout flow in the right state end to end, did the dashboard render the correct data, did the onboarding modal show up exactly when it was supposed to. The realistic stack is both, scoped to what each is good at. The win of the AI approach is that you stop maintaining __snapshots__ folders for tests that are really asking page-level questions.

Everything stays on your disk in standard formats. The scenario is a .md file (regex #?\s*(?:Scenario|Test|Case) for the case header). Results are JSON matching the TestReport shape at types.ts lines 28-35. Screenshots are PNGs, the video is a WebM, the player is a plain HTML file. There is no vendor baseline store, no proprietary diff format, no cloud-only dashboard. The comparable tier-3 AI testing platforms charge around $7,500 a month at scale. Assrt ships as an open-source npm package (npx assrt-mcp) you run against your own app. Cancel the relationship and the scenario still runs, the video still plays, the results still parse.