Get the latest on AI, LLMs & developer tools
New MCP servers, model updates, and guides like this one — delivered weekly.
What Watchr Is
Watchr is an MCP server, published as the watchr-mcp package, that turns Claude Code into a QA agent for iOS, Android, and web apps. Its own tagline is precise: “The AI research & QA agent for iOS, Android & web. Built on MCP for Claude Code.” You describe a testing task in plain English, and Watchr drives the real app, hunts for bugs, and hands back screenshots and a narrated report.
The useful way to think about it: Watchr is not a test framework and not a hosted cloud-testing grid. It is a tool layer that plugs an agent you already use, Claude Code, into real device and browser automation. The maker, Affinity Labs, describes it as one of five technologies in its internal “Foundry” engine, now published publicly on PyPI (v0.16.2 as of June 2026).
- What: MCP server (
watchr-mcp) that gives Claude Code QA superpowers. - Platforms: iOS Simulator, iOS device, Android, and web (Playwright).
- Style: plain-English instructions, no scripts or selectors; screenshots and a report by default.
- Edge: one surface across mobile and web; most agent-native QA tools are web-only.
Why It Exists
Test automation on mobile and web is slow to build and expensive to maintain. Appium, XCUITest, Espresso, and Selenium suites need specialist engineers and break whenever selectors or the DOM shift. Manual exploratory testing does not scale and is not repeatable across three platforms. Small teams shipping fast rarely have budget for a dedicated QA-automation function, yet still need to ship with confidence.
Watchr's wedge is to let a developer, or their agent, test in plain English across iOS, Android, and web, and get autonomous exploration plus a full QA battery without standing up scripts. Affinity Labs quantifies the internal payoff as “30–60% QA time saved on every build.” That is the vendor's own metric and is not independently verified, but it tracks with the category's core value proposition.
Install and First Run
Setup is a single MCP registration. The command below is taken verbatim from Watchr's site; uvx fetches the package on demand, so there is no separate install step.
# Add Watchr as an MCP server in Claude Code (from the official site):
claude mcp add --scope user watchr -- \
uvx --python 3.12 --upgrade watchr-mcp
# uvx downloads everything automatically. Requires Python >=3.11,<3.14.
# Then open Claude Code in any project and just describe what to test.If you have not wired up MCP servers in your agent before, our MCP setup walkthrough covers the general pattern, and Claude Context is a good example of another MCP server you can run alongside it.
One Surface, Four Backends
The same plain-English commands run everywhere. Watchr auto-detects what is connected and, in its words, speaks the right dialect for each target.
| Backend | Setup |
|---|---|
| iOS Simulator | Auto-detected |
| iOS physical device | One-time guided setup |
| Android emulator and device | Zero setup |
| Web | Headless Chromium via Playwright (device="web") |
The web backend runs on headless Chromium through Playwright, which puts Watchr in the same lineage as the Playwright tooling we covered in our stealth Chromium for Playwright guide. The difference is that here the agent, not a script, decides what to do.
What It Can Do
Watchr exposes a set of named MCP tools the agent calls. They split into control, quality checks, and reporting. A single session can combine many of them at once.
| Tool | Group | What it does |
|---|---|---|
explore | Control | Autonomously crawls an app from a starting screen, mapping flows and surfacing pages you forgot existed. |
parallel_run | Control | Fans out several sessions across several flows in one call. |
personas | Control | Behaves like a defined user, for example a first-time user on slow 3G or a power user with a full cart. |
frustration signals | Control | Instruments the page for rage clicks, dead clicks, and thrash, UX-research style. |
check_app_state | QA | Reads crash logs and console errors after a risky action. |
audit_accessibility | QA | Flags missing labels and too-small touch targets, per screen. |
perf_audit | QA | Flags slow (over 3s) and very slow (over 5s) screens. |
compare_with_baseline | QA | Visual regression: diffs against golden screenshots. |
monkey_test | QA | Chaos testing: random taps and swipes to shake out crashes. |
usability_audit | QA | Heuristic UX review of the current screen. |
security_scan / seo_audit | QA | Web-side security and SEO scans. |
assert_* | QA | Assertions: text visible, URL matches, no console errors, no failed requests. |
build_report | Reporting | Turns any session into a shareable QA report; can import test cases from CSV. |
The research-agent tools are what set it apart from a test runner. explore and personas mean it can investigate an app the way a curious tester would, not just replay recorded steps, and frustration signals borrow from UX research rather than pass/fail QA.
A Real Workflow
There is no DSL to learn. After the MCP server is registered, you prompt Claude Code in plain language and Watchr does the driving.
# After installing, you talk to Claude Code in plain English. Examples: "Test the checkout flow I just changed on iOS and tell me what broke." "Reproduce the login bug on Android and capture the network traffic." "Run an accessibility and performance audit on the settings screens." "Poke around the onboarding and flag anything that feels off." # One session can be a functional test, an a11y audit, a perf check, # a visual-regression pass, and a crash hunt at the same time. # You get screenshots at every step and a narrated report at the end.
Watchr lists four moments where it fits: testing the flow you just changed before a PR merges, a single multi-platform pre-release pass covering function, accessibility, performance, and crashes, reproducing and instrumenting a specific bug, and open-ended exploratory testing of an area that feels risky.
How It Compares
Watchr sits in a category the industry started naming in 2026: agent-native autonomous QA, meaning quality tooling that AI coding agents call directly through MCP and that runs and reports tests without step-by-step human instruction. Most of that tooling is web-only today, which is exactly where Watchr's mobile coverage stands out.
| Tool | Focus | Note |
|---|---|---|
| Watchr | iOS + Android + web, via MCP | Plain-English, with research-agent behaviors (exploration, personas, frustration signals). Cross-platform mobile is its least-crowded edge. |
| Playwright MCP (Microsoft) | Web | The most widely documented browser MCP for Claude Code; 2026 added self-healing test repair. |
| mobile-mcp and similar | iOS / Android | The most direct conceptual rivals for driving mobile apps through MCP. |
| Octomind, QA Wolf, Mabl, Reflect | Web-first | AI-generated and self-healing test platforms, often hosted SaaS rather than an agent tool layer. |
| Computer-use agents (Anthropic, OpenAI) | General | Can drive apps but are not QA-specialized; no built-in audits or regression tooling. |
The comparison points other than Watchr are general landscape context; their exact head-to-head behavior against Watchr is not vendor-confirmed. The reliable conclusion is narrow and useful: if you only test web, Playwright MCP is the well-trodden path; if you need one agent surface across iOS, Android, and web, Watchr is doing the less common thing.
Caveats and Unknowns
- Early version. At v0.16.2, this is pre-1.0 software; expect rough edges and change.
- Vendor metric. The “30–60% QA time saved” figure is Affinity Labs' own claim, not an independent benchmark.
- Commercial model is unclear. The package is free on PyPI and there is no pricing page; it is framed mostly as internal Foundry infrastructure as of June 2026.
- Company facts are thin. Watchr is built by Affinity Labs, reportedly an Antler-backed London startup (founders named in search as Ahmed Eltom, Edoardo Ricci, and Sunny Panchal). The landing page itself names no founders, funding, or pricing.
- Name collision. Do not confuse it with watchr.ai, an unrelated competitive-intelligence product that also ships an MCP server.
Use It If / Skip It If
Use it if you already drive development through Claude Code, ship across mobile and web, and want exploratory and regression QA without maintaining an Appium or Playwright suite. The plain-English surface and built-in audits are a real time saver for small teams with no dedicated QA-automation engineer.
Skip it, for now, if you need a stable, contractually supported QA platform with SLAs, if you are web-only and already happy with Playwright MCP, or if your org cannot adopt pre-1.0 tooling that runs through an external MCP package. Re-evaluate as it matures past v1.0 and the commercial model becomes clear.
FAQ
What is Watchr?
Watchr is an MCP server, published as the watchr-mcp package, that turns Claude Code into a QA agent for iOS, Android, and web apps. You describe a test in plain English; Watchr drives the app, runs a QA battery, and reports back with screenshots. No scripts or selectors.
How do I install Watchr in Claude Code?
Run: claude mcp add --scope user watchr -- uvx --python 3.12 --upgrade watchr-mcp. uvx downloads the package automatically. It requires Python 3.11 or later, below 3.14.
Does Watchr test mobile apps or just web?
Both. Watchr targets the iOS Simulator, iOS physical devices, Android emulators and devices, and the web through headless Chromium (Playwright). Unifying mobile and web behind one plain-English surface is its main differentiator.
Do I need to write test scripts or selectors?
No. Watchr's pitch is no scripts and no selectors. You give plain-English instructions and the agent figures out the taps, typing, and navigation, then reports what it found.
Is Watchr free?
The watchr-mcp package is published free on PyPI, and the site has no pricing. Watchr is framed primarily as internal infrastructure that Affinity Labs productized, so whether it becomes a paid product is unconfirmed as of June 2026.
Is Watchr the same as watchr.ai?
No. watchr.ai is an unrelated competitive-intelligence and brand-monitoring product that also happens to ship an MCP server. The Watchr in this article is the Affinity Labs QA agent, package watchr-mcp.
Sources
Product facts are from Watchr's own site and package; sector and company context is labeled as such.
Primary (Watchr and Affinity Labs)
- Watchr landing page
- watchr-mcp on PyPI (v0.16.2)
- Affinity Labs: The Foundry (Watchr is tech #03)
- Affinity Labs company site
- Affinity Labs GitHub org
Sector and market context