AI tools for testing is a list of 18 function calls in one TypeScript file.

At the level of the Assrt MCP server it means a fixed list of 18 function-call schemas that a Claude Haiku 4.5 instance is allowed to invoke. They are declared as a const named TOOLS at assrt-mcp/src/core/agent.ts line 16, and they are the only surface the agent is exposed to. navigate, snapshot, click, type_text, select_option, scroll, press_key, wait, screenshot, evaluate, create_temp_email, wait_for_verification_code, check_email_inbox, assert, complete_scenario, suggest_improvement, http_request, wait_for_stable. Clone the repo, grep for 'name: "' in that file, and you will count exactly 18. Nothing else happens server-side. No hidden cloud DSL, no proprietary selector engine.

Two tools do the work together. create_temp_email creates a real disposable address through temp-mail.io (the agent calls POST https://api.internal.temp-mail.io/api/v3/email/new at email.ts line 44), then wait_for_verification_code polls the inbox every 3 seconds for up to 60 by default, 120 seconds max. When mail arrives, a cascade of 7 regexes tries to extract the code, most specific first: code/Code/CODE, verification/Verification, OTP/otp, pin/PIN/Pin, then plain 6-digit, 4-digit, 8-digit runs. That cascade is at email.ts lines 101 to 109. You never touch the email yourself.

No, and the workaround is hard-coded into the agent system prompt at agent.ts line 235 so the LLM does not have to reinvent it per site. Instead of typing into each input, the agent calls the evaluate tool with a specific JavaScript expression that dispatches a synthetic ClipboardEvent carrying the full code to the parent of the first input. React and most OTP widgets treat a paste as one atomic fill, so all six boxes populate at once and the Verify button becomes clickable. The expression literally starts with `() => { const inp = document.querySelector('input[maxlength="1"]');`. You can read it verbatim in the repo.

Most AI test tools paper over streaming UIs with a fixed sleep. Assrt installs a MutationObserver on document.body via evaluate, stores the running mutation count at window.__assrt_mutations, then polls every 500ms and resolves as soon as the count stops increasing for a configurable quiet period (2 seconds by default). Implementation lives at agent.ts lines 956 to 1009. In practice this means an AI chat app that streams tokens for 14 seconds becomes testable: the agent waits exactly as long as the stream takes and no longer.

One TypeScript file. /Users/matthewdi/assrt-mcp/src/core/agent.ts is where the TOOLS array, the SYSTEM_PROMPT, the agent loop, wait_for_stable, and the OTP instructions live. email.ts is 131 lines and holds the disposable inbox plus the 7-regex cascade. browser.ts wraps @playwright/mcp. The published npm package @assrt-ai/assrt ships the same files under dist/, so `grep -n scenarioRegex node_modules/assrt-mcp/dist/core/agent.js` confirms the code you ran is the code you read. MIT license. Everything self-hostable.

Competitor action lists tend to be feature checkboxes across a vendor's entire product: reporting, dashboards, parallelisation, analytics. The 18 here are only the function-call schemas the LLM sees at inference time. Fewer, sharper tools mean the Haiku agent picks the right one on almost every step and rarely fabricates a non-existent one. Adding more would bloat the prompt, lower the pick accuracy, and push us toward tool-selection errors. The whole point of the abstraction is that your English is the grammar and a small, orthogonal palette is the verbs.

Most top-ranked options (Testim, Mabl, Functionize, Applitools, testRigor, Virtuoso, Checksum, ACCELQ) are cloud-hosted platforms with a proprietary scripting format, a recorder that captures a DOM path object repository, and monthly per-seat pricing in the $150 to $7,500 range. Your tests live on their servers. Assrt is an MIT-licensed Node package that dispatches to the official @playwright/mcp server. Tests are a Markdown file on your disk. The browser is the Chromium binary on your machine. There is no cloud dependency; app.assrt.ai is opt-in sync only.

The default is claude-haiku-4-5-20251001 (see DEFAULT_ANTHROPIC_MODEL at agent.ts line 9). The TestAgent constructor also accepts a Gemini provider and a default of gemini-3.1-pro-preview, plus an authType of 'oauth' for Claude Code sessions and 'apiKey' for regular API keys. Override with --model on the CLI or ANTHROPIC_MODEL / GEMINI_MODEL environment variables. Bring-your-own key; Assrt never proxies your traffic.

It falls back in the order prescribed by the SYSTEM_PROMPT at agent.ts line 220. First it calls snapshot again to get a fresh accessibility tree with ref IDs (the rule on line 213: 'ALWAYS call snapshot FIRST'). Then it tries a different ref or a scroll. If three attempts fail it calls assert with passed:false and an evidence string, captures a screenshot (emitted to the run report), then complete_scenario. You do not get a silent hang; you get a red result with a frame at the moment of failure saved under /tmp/assrt/<runId>/.

Yes. assrt_plan navigates to a URL, queues up to 20 pages (MAX_DISCOVERED_PAGES at agent.ts), and fans out up to 3 concurrent Haiku calls (MAX_CONCURRENT_DISCOVERIES) that each generate 1 to 2 #Case blocks per page from the live accessibility tree plus a screenshot. Feed that output straight into assrt_test. The tooling is intentionally small enough that the discovery pass and the execution pass share the same browser session and the same TOOLS palette.