How to do visual regression testing when the output is a folder, not a dashboard.

It means three separate install steps collapse into one. With a traditional setup you install a runner (Playwright, Cypress), an assertion library (toHaveScreenshot, pixelmatch, resemble.js, Percy SDK), and a hosting account for the baseline store (Percy cloud, Chromatic, Applitools Eyes). With Assrt, running `npx assrt setup` once registers the Assrt MCP server globally for Claude Code (scope: user), writes a PostToolUse hook to ~/.claude/settings.json, and appends a QA testing section to ~/.claude/CLAUDE.md. After that the `assrt_test` tool is always present in the agent's tool list. There is no per-project config file and no baseline folder. The scenario is a plain-text `.md`, and the result lives on your disk at /tmp/assrt/<runId>. Source: assrt-mcp/src/cli.ts lines 214-308.

Every run creates a freshly-minted UUID directory at /tmp/assrt/<runId>/ (assrt-mcp/src/mcp/server.ts line 430). Inside that folder you get: screenshots/NN_step<n>_<action>.png, one per visual action, with a zero-padded index starting at 00 (naming pattern at server.ts line 468); video/recording.webm, the native Playwright recording (server.ts lines 587-590); video/player.html, an auto-generated self-contained HTML video player (server.ts line 619 calls generateVideoPlayerHtml); execution.log, a flat text trace of every step and assertion (server.ts line 602); events.json, the full event timeline (server.ts line 606). Results land separately at /tmp/assrt/results/<runId>.json and /tmp/assrt/results/latest.json. None of these files are uploaded to a cloud by default. Open any folder in Finder and you have the full review surface.

It is 48 lines of HTML + CSS + vanilla JS, written by generateVideoPlayerHtml at assrt-mcp/src/cli.ts lines 310-349. It renders the .webm inline with autoplay, a dark chrome, a status header showing pass/fail counts and duration, and a row of playback speed buttons (1x, 2x, 3x, 5x, 10x). Keyboard shortcuts are bound: Space toggles play/pause, ArrowLeft seeks back 5 seconds, ArrowRight seeks forward 5 seconds, and keys 1/2/3/5/0 set speed to 1x/2x/3x/5x/10x. The file is self-contained, so you can zip the whole runId folder, mail it to a teammate, and they can double-click player.html on any machine. No proprietary player, no auth token, no browser extension required.

It uses Playwright directly. The Assrt browser layer spawns @playwright/mcp as a stdio child process with --viewport-size 1600x900 --caps devtools (assrt-mcp/src/core/browser.ts line 296). The native MCP tool calls the agent dispatches are real Playwright actions: browser_navigate, browser_click, browser_type, browser_take_screenshot, browser_start_video, browser_stop_video. The recorded WebM is Playwright's native video, not a synthetic reconstruction. Nothing in the pipeline is a proprietary YAML DSL. The plan format is plain English in scenario.md and the execution layer is Playwright MCP. If you cancel Assrt tomorrow, the plan still runs on any other Playwright MCP agent because the tool calls are public.

toHaveScreenshot() writes a golden PNG to __snapshots__/ on the first run, then pixelmatches every later run against that baseline. You tune maxDiffPixels and threshold to control false positives, and you run `npx playwright test --update-snapshots` any time you intentionally change the UI. Assrt has zero references to toHaveScreenshot, pixelmatch, resemble.js, or maxDiffPixels in the repo. Each visual action takes a JPEG screenshot, attaches it to the next Claude Haiku 4.5 tool-result message as base64 (agent.ts line 987), and the model decides whether the frame matches the English plan. Instead of a pixel count, the fail signal is an evidence string. Both approaches are valid. Pixel diffs catch 1px border drifts better. Semantic evidence catches page-level behavior better and does not need a __snapshots__ folder that churns on every theme tweak.

Three commands cover 95% of cases. First, open the auto-generated HTML player: `open /tmp/assrt/<runId>/video/player.html` (or just let assrt_test open it for you, autoOpenPlayer defaults to true, server.ts line 403). Second, read the assertion list: `jq '.scenarios[].assertions' /tmp/assrt/results/<runId>.json` gives you `description`, `passed`, `evidence` for every check (shape at assrt-mcp/src/core/types.ts lines 13-17). Third, look at the exact frame that failed: `open /tmp/assrt/<runId>/screenshots/` lists them in order because the filename prefix is a zero-padded numeric index. No dashboard login, no signed URL expiry, no 'session has ended, please refresh'.

Zip the /tmp/assrt/<runId> folder and upload it. Everything is already a standard format: PNG, WebM, HTML, JSON, plain text. On GitHub Actions: `- uses: actions/upload-artifact@v4` pointing at `/tmp/assrt/<runId>`. On GitLab CI: add it to `artifacts.paths`. The HTML player is self-contained so reviewers can unzip the artifact locally and double-click the player to replay the run, no GitLab Pages deploy required. Because the scenario.md is also plain text you can commit it to your repo next to the code it tests and diff it in PRs, which is painful with most proprietary scenario formats.

Pass `viewport` to assrt_test as either a preset string or an explicit object like `{ width: 414, height: 896 }` (TestRunOptions.viewport at assrt-mcp/src/core/types.ts line 46). Each run creates its own /tmp/assrt/<runId>/ so a mobile run and a desktop run against the same scenario produce two independently reviewable folders. Playwright's default browser is Chromium and Assrt uses the bundled chromium binary from @playwright/mcp. If you need Firefox or WebKit, swap the channel in the browser launch config at browser.ts line 296. Viewport defaults to 1600x900 for the MCP spawn (browser.ts line 296) and the video recording honors the same dimensions via browser_start_video at line 628.

Optional. The default mode writes every artifact to your local disk and makes zero outbound calls beyond the LLM request for semantic judgment (Anthropic for Claude Haiku 4.5, optionally Google for Gemini 3.1 Flash Lite retrospective video analysis if GEMINI_API_KEY is set). If you sign in with `assrt login` the server also mirrors scenario metadata to Firestore for multi-device sync, but that path is guarded by ASSRT_NO_SAVE=1 (server.ts line 404). The comparable tier-3 AI testing platforms charge roughly $7,500 a month at scale and keep scenarios, diffs, and evidence in their cloud permanently. Assrt ships as an open-source npm package @assrt-ai/assrt and a marker file at ~/.assrt/installed confirms an idempotent install. Cancel the relationship and every past /tmp/assrt/<runId> folder still plays, parses, and reproduces.

The same `assert` tool handles them. Its three fields (description, passed, evidence) are model-agnostic, so the agent can make an http_request, inspect the response body, and emit an assertion with evidence like 'POST /api/orders returned 201 with order_id ord_8821'. The http_request tool at agent.ts lines 172-184 supports GET/POST/PUT/DELETE and headers. A single scenario can mix visual assertions and integration assertions, which matters when you are testing an end-to-end user journey: the toast says 'Order placed' visually and the backend has a new row in the orders table. Built-in means the same plan format covers both.