A visual regression framework is six files on your disk, not a dashboard you rent.

A framework is a set of primitives you compose to solve a problem class, not a single tool. Playwright is a framework; Chromatic is a platform. Most 'visual regression frameworks' the SERP recommends are platforms: they own the dashboard, the baseline store, the diff algorithm, and the access control. A real framework publishes its primitives as files you can inspect. Assrt's primitives are the plan file (scenario.md), the metadata file (scenario.json), two flavors of the results file (results/latest.json and results/<runId>.json), the per-run screenshot directory, and a WebM video. Six paths. Everything else is commentary.

All under /tmp/assrt. The layout is documented at the top of assrt-mcp/src/core/scenario-files.ts lines 5 through 20. The plan lands at /tmp/assrt/scenario.md. Metadata (id, name, url, updatedAt) lands at /tmp/assrt/scenario.json. The most recent run's results land at /tmp/assrt/results/latest.json, and every historical run is archived at /tmp/assrt/results/<runId>.json. Screenshots for a run are at /tmp/assrt/<runId>/screenshots, and the WebM video plus a standalone player.html live at /tmp/assrt/videos (set in assrt-mcp/src/mcp/server.ts around line 618). If you are testing CI-side, point these paths at a persistent volume or artifact store and you have a full archive of every run.

Because the agent reads it, humans read it, and git reviews it. A .md file is the least-proprietary text format alive. The parser in assrt-mcp/src/core/scenario-files.ts is effectively a single regex on case headers (#?\s*(?:Scenario|Test|Case)), so Scenario, Test, and Case all work. The body is imperative English. No schema to validate, no version field to migrate, no Chromatic-specific keys. You can copy-paste a scenario.md into any other Playwright MCP agent and it runs unchanged, because the plan says 'go to /dashboard, verify the sidebar avatar is a real image' and every MCP agent understands that sentence.

It is a TestReport at assrt-mcp/src/core/types.ts lines 28 through 35. Fields: url, scenarios, totalDuration, passedCount, failedCount, generatedAt. Each scenario is a ScenarioResult (lines 19 through 26) with name, passed, steps, assertions, summary, duration. Each assertion is three fields: description, passed, evidence (lines 13 through 17). No proprietary IDs, no signed URLs that expire in seven days. You can grep failedCount across a year of archived run files and answer real questions like 'what is my visual regression fail rate by week'.

That tutorial focuses on why pass/fail uses an LLM reading a JPEG instead of pixelmatch diffing a baseline PNG. This page is the framework layer above that: what files are produced, what formats they are in, and why that matters if you ever want to leave your vendor. If you are deciding whether to adopt Assrt, read the tutorial for the reasoning model and read this for the data-portability model. They answer different questions. Both are grounded in the same repo.

Yes. /tmp/assrt is the default because /tmp is writable on every CI image without extra config. In GitHub Actions you can copy the directory as a build artifact after the run, or symlink /tmp/assrt to a mounted volume before the run. The paths themselves are stable: scenario.md and scenario.json always sit at the root of the directory, results land under results/, screenshots under <runId>/screenshots, and the video under videos/<runId>.webm. None of those paths contain random UUIDs you have to discover. If you want the raw list of written files, assrt-mcp/src/mcp/server.ts lines 431 through 486 and lines 618 through 629 are the authoritative write sites.

Nothing. They are on your disk. npx assrt-mcp is the open source npm package that wrote them, and you can keep running it offline against any Playwright MCP-compatible agent. The plan is English; any model that speaks Playwright MCP executes it. The results JSON is plain, unencrypted, schema-documented TypeScript types. The screenshots are PNGs. The video is a WebM playable in every modern browser with Range support. The comparable tier-3 AI visual regression platforms charge around 7,500 dollars a month and keep your scenarios, selectors, and diff history in their cloud; cancel there and the login page is the last thing you see. On Assrt, cancel and tar /tmp/assrt; the archive replays.

There is an optional hosted view at assrt.ai for teams that want one, but it is not the framework. The framework is the local files, exposed through the MCP tools assrt_test, assrt_plan, and assrt_diagnose. If you run everything locally and never log in to the cloud view, every feature still works because the agent reads and writes the local files directly. The cloud view is a mirror, not the source of truth. That inversion is the whole point of calling this a framework instead of a platform.

Pass criteria live inline in scenario.md under a 'Pass Criteria' sub-section of a #Case; the agent reads them as additional assertion targets for that case. Screenshots-per-step are automatic: after any visual action (navigate, click, type_text, select_option, scroll, press_key) the agent captures a JPEG, emits it to the model for reasoning, and also writes it to disk at /tmp/assrt/<runId>/screenshots/<index>_step<stepNumber>_<action>.png (naming convention at assrt-mcp/src/mcp/server.ts line 468). The deduplication logic at server.ts lines 473 through 486 keeps only the last frame per step, so a run with 12 visual actions produces at most 12 PNGs.

No. Open source is necessary but not sufficient. Many OSS visual regression frameworks still assume you run their central diff service, and the scenario or baseline format is theirs. Lock-in survives the license if you cannot move the artifacts. Assrt's stronger claim is that every artifact it writes is either a public file format (.md, .json, .png, .webm) or a documented TypeScript type (types.ts lines 13 through 35). There is no Assrt-proprietary DSL. You can point a different Playwright MCP agent at the same scenario.md and run it with zero edits, because the tools the agent uses (navigate, click, type_text, and the rest) are the public Playwright MCP vocabulary, not Assrt's invention.