QA Reef vs QA Wolf
QA Wolf is the platform most people mean when they say "AI QA testing." Every claim about QA Wolf below is sourced — docs.qawolf.com, qawolf.com, or a named third party — and every claim about QA Reef is limited to what's actually implemented, not planned.
Researched 2026-08-28 from public sources; verify current details at qawolf.com.
The short version
Both tools follow the same shape — map a flow, generate a test, run it, triage failures. The differences are underneath: what's open, screen vs. DOM, who verifies a failure, and what it costs.
| Dimension | QA Reef | QA Wolf |
|---|---|---|
| License / openness | Automation core (flow runtime, operator, OCR/pixel measurement, LLM adapter, evidence storage) is MIT, built in the open under packages/core — see Open source. |
Bottom layer only: Playwright (Apache-2.0) + Appium, plus QA Wolf's own @qawolf/flows (MIT) and @qawolf/cli (Apache-2.0), are public. Mapping AI, Automation AI, the runner grid, and the human QA service are closed. Source: github.com/qawolf. |
| How it finds things to click | OCR-driven operator reads the screen (Apple Vision or a vision-model adapter); a cheap LLM picks the next visible label to click. Generated tests still use ordinary Playwright locators once recorded. | DOM-based Playwright for recorded/generated flows. Its separate "Runner" (documented in the CLI's Agent Skill, not the public docs) is a coordinate-based act/screenshot loop, screenshot-on-demand, no continuously-running hosted vision model — an external agent decides each action. Source: @qawolf/cli npm package, skills/qawolf-cli/references/runner.md. |
| Self-healing / diagnosis | Not implemented yet. Recorded steps become Playwright locators at generation time; there is no automatic re-resolution of a broken locator today. | DOM-based: failure analysis reviews HTML at error time, visible code, page title/URL, and action history; self-healing evaluates DOM diffs, network traces, console errors, and fixture state — not pixels. Source: qawolf.com/blog, qawolf.com/blog. |
| Model-decision transparency | Every agent click, type, or assertion is logged with the exact model id (decided_by) and token cost, visible in the product. |
Not publicly documented at this level of detail; billing is by "AI credit," not per-call model attribution. Source: qawolf.com/pricing. |
| Verdict states | PASS / FAIL / UNMEASURED — a run that couldn't get a clean look is never silently counted as a pass or a fail. See why a QA verdict needs a third state. | Not a distinct concept; failures are bucketed flake / bug / broken-test, and a QA engineer manually verifies before a bug is reported — the "zero flake" claim is human-verification, not automated measurement. Source: qawolf.com/service, docs.qawolf.com. |
| Human-verified failures / managed maintenance | Not offered. Bugs are auto-filed from failed runs; a human on your team still triages them. | Yes — "Coverage as a Service" pairs the platform with dedicated QA engineers who reproduce and maintain tests; QA Wolf's own moat, by their description. Source: qawolf.com/service. |
| Mobile testing | Not supported. Playwright/Chromium only today. | Yes, via Appium. Source: github.com/qawolf, qawolf.com/blog. |
| Hosted execution grid | Not offered yet. Run locally, in your own CI, or self-hosted. | Pre-warmed Kubernetes runners, one container per test, dynamic autoscaling. Source: qawolf.com/blog. |
| Pricing model | Not published yet — see pricing page. Self-hosted core costs whatever your own model calls cost, and every call is cost-logged. | Self-serve: 1¢ per AI credit (Mapping/Automation AI) plus 15¢ per runner-minute, no seat fee, unlimited parallel runs. Coverage as a Service: not publicly priced, billed by tests under management. Source: qawolf.com/pricing (accessed 2026-08-28). |
On pricing, specifically
QA Wolf's self-serve rate card is 1¢ per "AI credit" plus 15¢ per runner-minute, no seat fee (qawolf.com/pricing, accessed 2026-08-28). Its main product, Coverage as a Service, isn't publicly priced — third-party (not QA Wolf's own) estimates put it around $40–70/test/month, ~$90k median annual contract, from ~$8k/month for 200 tests; unverified, directional only. Source: sacra.com/c/qa-wolf.
QA Reef has no published price yet — a self-hosted core only costs you the model calls it makes, and every call is logged. See the pricing page, or the full sourced QA Wolf pricing breakdown.
Go deeper
QA Wolf alternative
Fit, verifiable rows only, and how migrating a flow works.
QA Wolf vs QA Reef, in depth
Architecture, evidence and verdicts, hosting, and a sourced review of QA Wolf.
Open source QA Wolf alternative
What's in packages/core, the flow converter, and self-hosting.
Background, for context
- Founded 2019, Seattle, by Jon Perl and Laura Cressman; exited stealth Sept. 2022 (TechCrunch, 2022).
- $20M Series A (Inspired Capital, 2022), $36M Series B (Scale Venture Partners) (TechCrunch, 2024).
- Third-party estimate: ~$15–20M ARR, ~130 customers, unverified (sacra.com/c/qa-wolf).
QA Reef is pre-launch: no customers, no funding announcement, no case studies.
Want the architecture detail behind this table?
Engineering notes on QA Wolf's Runner, recorder, and self-healing — sourced from their own docs and CLI package — are on the blog.
Read: why we read the screen with OCR instead of trusting selectors →