Describe the behavior. Argus watches, acts, and proves it.
Argus tests desktop apps, web UIs and command-line tools the way a person would, using multimodal models. It looks at the screen, decides what to do, acts, and checks the result. Run it locally for quick checks, or inside a disposable Capsule when your own desktop shouldn't be part of the experiment.
Latest release →From intent to evidence, one boundary at a time.
Argus keeps placement, interaction, policy and verification separate. The model proposes; Argus checks every action before it runs. Assertions are never left to the model's opinion.
Observe → propose → validate → act
A hundred eyes on your app. None on your desktop.
Adapters decide how Argus interacts with a target. Execution environments decide where it's allowed to. Pick the boundary that fits the risk.
Capsule
Recommended for productionOne test style. Three ways in.
Tests are YAML files in .argus/. Plain-English steps go to the model; assert blocks run deterministically.
The model handles the judgment. Argus holds the guardrails.
04 — CapabilitiesNatural-language actions
Say what a user should do instead of maintaining brittle selectors for every interaction.
Deterministic assertions
Visible text, process state, exit code, URL and title are checked without asking the model.
Central execution policy
Every action is normalized and checked against capabilities and host-safety policy first.
Safe local mode
Supported desktop actions use semantic UI Automation, not host-wide input injection.
Disposable Capsules
Invasive tests run in dedicated VMs with secure guest control and explicit lifecycle ownership.
Failure Capsules
Keep disk and config evidence from a failed scripted run, without persisting live credentials.
Explicit file transfer
Stage inputs and collect outputs through declared, bounded, hashed transfers.
ATES v0.1 evidence
Ordered events, protected artifacts and verifiable manifests. Reports derive from evidence.
Any provider
Ollama locally, or Anthropic, OpenAI, Azure, Gemini and LiteLLM. Vision is auto-detected.
Let it wander. It writes down what breaks.
Give Argus a target and a time budget. It explores without a script, catches crashes and error dialogs, records findings with severity, and leaves you a report plus regression-test stubs. Cross-session memory means each session starts where the last one stopped.
$ argus roam "http://localhost:3000" --adapter browser
$ argus roam "./my-cli.sh" --adapter cli
What's shipped, and what's next. Stated plainly.
Execution authority, evidence authority and future read-only monitoring stay separate trust boundaries.
ATES v0.1
Canonical ordered evidence, durable action dispatch, privacy controls, protected artifacts, transactional finalization, manifests and verification, traceability, approvals and audit.
Read ATES →Fleet execution plane
Node identity and enrollment, authenticated heartbeats, placement and fencing, remote Capsule execution, and ATES transport above the existing Capsule boundary.
Read Fleet →Fleet operations + Observer
Scheduler and queues, matrix execution, timelines and indexes, and a strictly read-only Observer. Not presented as shipped.
Roadmap →Cross-platform releases
CI builds Windows x64 and ARM64, macOS universal2, and Linux x86_64 and ARM64 packages with checksums and provenance.
Latest release →Desktop app or Python package.
Every release builds native packages from the same tagged source, with SHA-256 checksums and GitHub build-provenance attestations. Needs Python 3.10+ for the package route; install Ollama for local models.