The natural language test case descriptions tool is the three capabilities it ships, not the marketing copy around the English input.

A tool that accepts English descriptions of test cases and turns them into something runnable. In the top SERP results this almost always means an English-to-test-script compiler: you write a sentence, the tool emits Selenium/Playwright code, and that code runs later. Assrt is in a different subcategory: the English is the program. The agent reads your #Case block and the live ARIA tree on every turn and picks one of 18 tool calls. Nothing is compiled in advance.

Three runtime capabilities that those tools either skip or paywall. First, a built-in disposable inbox: `create_temp_email` + `wait_for_verification_code` + `check_email_inbox` call temp-mail.io directly, so your signup #Case works end-to-end without a hard-coded test address. Second, `wait_for_stable`: the agent injects a MutationObserver into the live page and only returns once the DOM has been quiet for N seconds. This is critical for chat, streaming AI responses, and lazy lists. Third, scenario continuity: one browser process per plan. `#Case 2` inherits cookies and localStorage from `#Case 1`, and the model sees the previous case's summary in its context.

Look at /Users/matthewdi/assrt-mcp/src/core/agent.ts, lines 956-1009. When the agent executes `wait_for_stable`, it calls `browser.evaluate(...)` to install `window.__assrt_observer = new MutationObserver(...)` and `observer.observe(document.body, {childList, subtree, characterData})`. It then polls `window.__assrt_mutations` every 500ms. When the counter stops incrementing for `stable_seconds`, it returns. Then it calls `observer.disconnect()` and deletes both globals. That is the real primitive: an observer injected into the target page, not a fixed sleep.

`DisposableEmail.create()` in email.ts:43-55 POSTs to `https://api.internal.temp-mail.io/api/v3/email/new` with `{min_name_length: 10, max_name_length: 10}` and stores the returned email + token. `waitForVerificationCode()` polls `/email/{address}/messages` every 3 seconds (default) and extracts the numeric code using a cascading regex. In your #Case you just write 'Get a disposable email' as an English sentence; the agent turns it into a `create_temp_email` tool call and the address is injected into the agent's prompt so the next `type_text` call uses it.

Because most real flows have a prerequisite. If you want to test 'post a comment', you need to be signed in first. In tools that run each test case in a fresh browser, you end up either (a) duplicating the login sentences at the top of every case or (b) extracting a shared setup hook in a config file — at which point the 'natural language' promise has leaked back into scaffolding. Assrt keeps the whole plan in one browser: `#Case 1: Sign up` leaves `#Case 2: Post a comment` already authenticated. The outer loop in agent.ts:462-493 never closes the browser between cases, and `previousSummaries` in runScenario feeds every prior result back into the next model turn.

There's an escape hatch: the `evaluate` tool runs arbitrary JavaScript in the page. That covers anything you would otherwise reach for with devtools — reading localStorage, triggering a custom event, posting to a postMessage channel. But the happier pattern is to write the sentence so it binds to one of the observation-grade tools (assert against visible text, use http_request to hit a webhook endpoint). If a sentence can't bind to any tool, the agent will usually call `suggest_improvement` with a rewrite hint instead of confabulating a pass.

Two options. Per run, pass a `variables` map with your plan: any `{{VAR_NAME}}` in the English text gets substituted before parsing (agent.ts:377-381). Long term, keep secrets in macOS keychain and let the MCP server read them via the keychain helpers in assrt-mcp/src/core/keychain.ts. Either way the sentence reads naturally: 'Type {{TEST_USER}} into the Email field.' The agent sees the substituted value; your markdown stays safe to commit.

Yes. Both assrt (the web app) and assrt-mcp (the MCP server / CLI) are Node packages with MIT licenses. `npx assrt-mcp` runs the MCP server; `npx assrt run --url <u> --plan "$(cat scenario.md)"` runs a plan from the CLI. Bring your own Anthropic or Gemini key. Artifacts land in /tmp/assrt/ by default: scenario.md, results/latest.json, and per-plan video recordings if you pass --video.