QA automation services, itemized: 18 browser primitives you are paying a retainer for

A human who composes browser primitives into scenarios. The primitives are the same on every retainer: navigate, click, type, select, wait, snapshot/read, assert, screenshot. Different DSLs (Cypress commands, Playwright Locator calls, Selenium WebDriver, proprietary YAML) dress them up differently, but the underlying unit of work is identical. Assrt exposes its set as 18 MCP tool handlers in /Users/matthewdi/assrt-mcp/src/core/agent.ts between lines 16 and 196. They are: 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. When a services firm invoices you for a QA automation engineer, the deliverable is a composition of those 18 operations, padded with a dashboard and a markup.

create_temp_email (line 115), wait_for_verification_code (line 120), check_email_inbox (line 128), and http_request (line 172) let the test agent itself complete a signup flow end to end. When the system prompt at agent.ts:228-236 detects a signup form, it tells the model to (1) create a disposable email, (2) submit the form, (3) poll the disposable inbox for the OTP, (4) paste the digits back into the form. No QA engineer is sitting by a mailbox. On closed AI QA platforms this step is usually skipped (the vendor asks you for a pre-provisioned test account) or priced as a separate add-on. On a services retainer it is a human cost. In agent.ts it is four tools and about 150 lines of glue in /Users/matthewdi/assrt-mcp/src/core/email.ts.

Two things. First, every primitive lives in a readable TypeScript array (the TOOLS constant at agent.ts:16-196), so when a test misbehaves you can grep the handler name and see exactly what the agent issued to the browser. Second, the primitives map directly onto Playwright MCP's own tool names. The browser manager at browser.ts:284 resolves @playwright/mcp/cli.js and spawns it over stdio; click becomes a Playwright MCP click, snapshot becomes the accessibility-tree snapshot Playwright already exposes. There is no transpilation step from 'Assrt DSL' into browser operations, because there is no Assrt DSL. A vendor YAML has to be compiled by the vendor at runtime, which is the layer that breaks when your app changes.

wait_for_stable polls the DOM and resolves only when no mutations have occurred for a configurable stable window (default 2 seconds, max 30 seconds of wall-clock wait). The system prompt at agent.ts:249-254 instructs the model to call wait_for_stable after any async action: submitting forms, sending chat messages, triggering search. In a consultancy workflow, the corresponding solution is 'retry three times with exponential backoff' plus a human watching the flake rate. wait_for_stable replaces both by binding the wait to the actual signal (DOM quiescence) rather than a timer guess. This is why the same #Case suite can run against a slow staging server and a fast local dev server without per-environment tuning.

It applies even more. In extension mode (browser.ts:258, arg --extension) the runner attaches to your already-running Chrome via Chrome DevTools Protocol instead of launching a fresh headless browser. That means cookies, logged-in state, and the browser fingerprint your users actually have. A services firm would script around your auth and require test credentials; the sharedBrowser in extension mode uses your credentials because it is your browser. The handshake is gated on a one-time Playwright MCP extension token (saved to ~/.assrt/extension-token by browser.ts:231-241), so subsequent runs are silent. This covers the 'test on prod without creating a new account every time' case most retainer contracts refuse to touch because of compliance handwaving.

Four shell lines. npx @assrt-ai/assrt setup registers the MCP server with Claude Code, Cursor, Zed, and Windsurf (sets up ~/.cursor/mcp.json, ~/.claude/mcp.json). Start your dev server. Ask your coding agent 'Use assrt_plan on http://localhost:3000' to auto-generate 5 to 8 #Case blocks into /tmp/assrt/scenario.md. Ask 'Use assrt_test' to execute them. Under the hood, assrt_test spawns @playwright/mcp, the agent loop picks tools from the 18-entry TOOLS array, Haiku reasons over each step, and you get pass/fail plus a video player on a local port. No account, no dashboard, no per-seat license.

2026 retainer ranges, from published sources: QA agencies average around $180K/year, in-house teams $312K/year, closed AI tools $12K to $24K/year, with hidden test-maintenance fees typically 20 to 50 percent on top. Assrt charges zero platform fee. The variable cost is Anthropic tokens at the claude-haiku-4-5-20251001 rate (agent.ts:9 sets this as DEFAULT_ANTHROPIC_MODEL). A 5-case, 3-to-4-step-per-case suite uses roughly 8k to 15k input and 2k output tokens, which at Haiku pricing clears well under a penny per run. You can also set ANTHROPIC_BASE_URL and route to a local LLM, in which case the platform fee stays zero and the variable fee is your GPU time.

Yes. Every primitive in TOOLS (agent.ts:16-196) is a thin adapter over a Playwright MCP tool. If you stand up your own @playwright/mcp session (npx @playwright/mcp@latest --extension or headless), you can issue the same tool names (browser_navigate, browser_click, browser_type, browser_snapshot) and drive the same Chromium. Your #Case plans in /tmp/assrt/scenario.md are Markdown; pasting them into a different LLM-driven runner works. There is no vendor format to convert. This is what 'zero vendor lock-in' means in code: the primitives are standard MCP calls, the plan format is Markdown, the runner is open source (MIT). The exit cost is a git add and a different npx binary.

assrt_diagnose (in server.ts:866, separate from the primitives) takes the failing #Case text, the error evidence, and the URL, and runs one Haiku call with a diagnose system prompt (server.ts:240-268). The model emits a Root Cause, an Analysis, a Recommended Fix, and a Corrected Test Scenario in the same #Case format so you can paste the fix straight back into /tmp/assrt/scenario.md. This is the QA-engineer-triages-a-red-build role in a retainer. It costs fractions of a cent, closes in seconds, and you can re-run it with a different model via the model parameter. The whole diagnose handler is about 60 lines because, again, it is one model call, not a ReAct loop.

Because the LLM treats the tools as a flat namespace. Anthropic's tool-calling API takes an array of Tool definitions and chooses among them by name; Gemini's is the same shape. The TOOLS constant at agent.ts:16-196 is intentionally a flat array so the schema the model sees is identical to the schema the runner dispatches against. If you want to extend Assrt with your own primitive (say, a Stripe-test-card helper or a headless-browser-console-log reader) you append one entry to TOOLS and add a handler in the dispatch loop. That is the full extension protocol. No plugin registry, no config, no yaml.