From agent observations to QA evidence.
A local QA runtime that gives coding agents browser actions, goal coverage, and structured symptom checks. The host model drives exploration; Argus keeps the evidence and the limits of each finding inspectable.
THE QA WORKFLOW
The difficult part is connecting an exploratory pass to evidence someone else can inspect. Each goal has an evidence window; reports retain exercised, blocked, and unfinished work instead of presenting an incomplete pass as comprehensive.
01
The host agent chooses goals and constraints. Argus keeps a coverage ledger visible throughout the session.
02
Semantic browser actions collect observations, screenshots, network signals, and a value-redacted action trail.
03
Structured text, HTTP, or replay checks return reproduced, not reproduced, or inconclusive for a specified symptom.
04
Reports connect findings to evidence and unfinished coverage. Built-in text and HTTP checks can run later without a model.
ENGINEERING DECISIONS
A modular monolith separates protocol, application, domain, and adapters. MCP and CLI entry points share the local runtime.
One process keeps deployment simple, but shares a failure domain. Concurrent sessions need separate runtimes; this is not a hosted multi-user service.
One operation lock orders browser actions and lifecycle changes. Queue and execution waits are bounded; cancellation closes the backend.
A slow action blocks later work. A timed-out click may already have changed the app, so side effects are never automatically retried.
A read-only loopback monitor uses SSE and a shared event ring. Sequence gaps trigger a refresh from the current snapshot.
Slow clients cannot create unbounded queues. Events expire and disappear on restart; the monitor is operational visibility, not a durable audit log.
Executable input values stay in a session vault. Exported replays mark required inputs instead of treating redacted placeholders as real values.
Exports need fresh input binding to replay. Screenshots can still contain application data, and saved state capsules can contain credentials.
The 0.7 implementation ledger records scripted BuggyTasks and DarkShop checks. This measures deterministic fixture coverage, not an agent's autonomous discovery rate.
The release ledger records checks for runtime boundaries, cancellation, storage, stdio, Chromium, and monitor behavior. CI is configured for Python 3.10 and 3.12 with Chromium.
No paid model benchmark was run for 0.7, and live native macOS UI was not exercised in that release validation. These results do not establish reliable autonomous QA across arbitrary applications.
QUICK START
With uv and Claude Code installed, Argus runs locally over stdio. The default profile keeps the MCP tool surface focused; no hosted service or Argus account is required.
$ 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
Browser QA, goal coverage, evidence capture, symptom verification, and reporting. The default MCP profile.
Native macOS actions through Accessibility and screenshots. Live native UI was not exercised in the 0.7 release validation.
The combined surface plus specialist controls for storage, network mocking, tabs, coordinates, crawling, and fault injection.
Profile counts checked against the October 19, 2026 source snapshot. Reports export as HTML, JSON, JUnit, and SARIF.
FAQ
The host model chooses the journeys and interprets product behavior. Argus supplies the actions, observations, coverage ledger, and verification machinery. Discovery quality still depends on the model, prompt, budget, and application.
It establishes that the specified observable predicate held under the check conditions. It does not establish root cause, product correctness, or complete test coverage. Text and HTTP verification reload twice with mocks suspended; authentication cookies remain, and clearing client storage is opt-in.
Automatic regression accepts built-in text and HTTP checks and refuses replay and custom strategies. Explicit replay uses a separate browser context, but can still repeat real backend writes. A passing regression run means no known symptom reproduced, not that the application is bug-free.
OPEN SOURCE · LOCAL RUNTIME · INSPECTABLE EVIDENCE