CORE29 tools
The focused default for exploratory browser QA, evidence capture, clean-state verification, and reporting.
OPEN-SOURCE · MCP · QA
Argus.
A testing engineer for your AI coding agent.
Argus is an MCP server that tests web and native macOS apps like a real QA engineer: it explores user journeys, discovers bugs nobody scripted a test for, and proves each certified finding before reporting it.
REPRODUCTION RECEIPTVERIFIED
✓
Profile update does not persist
The success toast appears, but a clean reload restores the previous display name.
- OBSERVED
- success feedback
- CLEAN RELOAD
- old value returned
- EVIDENCE
- before + after
- CI REPLAY
- model-free
A model's claim is not evidence. The clean-state re-check is.
THE PROOF LOOP
Discover first.
Certify second.
Existing end-to-end tests confirm the cases somebody already anticipated. Argus gives an agent the observation, action, and evidence tools needed to look for failures outside that script.
01
Explore
Follow real user journeys with semantic browser actions instead of a predefined assertion list.
02
Observe
Correlate visible behavior with layout, accessibility, console, network, persistence, and state evidence.
03
Reproduce
Reload from a clean state and independently re-check the symptom before certifying a finding.
04
Guard
Emit a replayable receipt and machine-readable reports so the bug can become a model-free CI gate.
BUGS BETWEEN THE ASSERTIONS
The page can look fine
and still be lying.
Argus is designed for behavioral, state, responsive, and UX defects that are difficult to catch with static analysis or a happy-path test suite alone.
- A “Saved” toast appears, but the change disappears after reload.
- Delete looks successful until the supposedly removed record comes back.
- A sale badge claims a discount while both prices are identical.
- Account data updates on one page but remains stale in the navigation.
- A responsive layout clips controls or leaves touch targets too small.
- A user journey fails through console, HTTP, link, or accessibility errors.
QUICK START
Two commands.
Then test the app.
Argus runs locally over stdio. The default profile keeps the MCP tool surface focused; no hosted service or Argus account is required.
TERMINALCLAUDE CODE · UVX
$ uvx --from playwright playwright install chromium
$ claude mcp add argus -- uvx --from argus-testing argus-mcpTHEN ASK YOUR AGENT
“Test my app at http://localhost:3000 — find real bugs.”
MATCH THE TOOL SURFACE TO THE JOB
SCREEN14 tools
Native macOS testing through Accessibility and screenshots, with action trails, timeouts, and a panic file.
FULL76 tools
Advanced browser and screen controls for network mocking, storage, tabs, coordinates, crawling, and fault injection.
- deterministic capability ceiling
- 34 / 34
- report formats
- 4
- independent reload verification
- CLEAN
- model calls during CI replay
- ZERO
34/34 measures what the deterministic benchmark fixtures expose through the tool surface. It is a capability ceiling, not a claim that every live model finds every bug in one pass.
HONEST BOUNDARIES
Proof improves trust.
It does not make QA exhaustive.
Discovery quality depends on the host model and the time and context it spends exploring. A weaker pass may find fewer bugs; the receipt controls what may be labeled verified.
Argus is black-box QA, not a replacement for unit tests, source review, security testing, or deliberate product-specific test design. Its strongest role is finding and proving unexpected behavior, then preserving it as a regression check.
FAQ
What Argus is—and isn't.
Is Argus a Playwright replacement?
No. Playwright is excellent for scripted automation. Argus uses Playwright as part of a broader exploratory QA workflow: it helps an MCP agent discover unknown bugs, prove them, and turn them into replayable regression evidence.
Does Argus require access to my source code?
No. Its primary workflow is black-box testing against a running app. Give the agent a URL; Argus observes and interacts with the product the way a user would.
Does a verified finding mean Argus found every bug?
No. Verified means the reported symptom survived an independent clean-state check. Discovery recall still depends on the host model, prompt, time budget, and product surface, so Argus reports its benchmark ceiling separately from live-model recall.
Can Argus test desktop apps?
Yes, on macOS. The screen profile drives native apps through the Accessibility tree and screenshots, with coordinate fallbacks for interfaces that expose limited accessibility data.
LET THE AGENT EXPLORE. MAKE THE EVIDENCE DECIDE.