Strix

Why it's ranked here

The open part is real software, not a teaser. It ships as an Apache 2.0 Python package with a console command, a strict type-checking and security-linting setup in its manifest, and a sandbox image that documents about thirty named security tools rather than leaving the agent to improvise. It also fits into CI with a documented exit code for findings and a fast scan mode for pull requests. Against that, the package declares itself Alpha, and several headline features, such as one-click autofix and continuous scanning, are described as belonging to the paid cloud product.

What's good

The toolbox is concrete: the docs list Nmap, Nuclei, SQLMap, ZAP, Semgrep, TruffleHog, Gitleaks, Trivy and Caido, each with its job, so you can reason about what the agent is able to do. Three scan depths are spelled out with time costs, from minutes for quick checks to one to four hours for the default deep mode. Headless runs print findings live and exit with code 2 when something is found, which is what a pipeline needs. Interrupted runs can be resumed by name, and the error handling names the exact optional extra to install when Bedrock or Vertex support is missing.

Tradeoffs

The package classifier says Alpha. Every scan spends model tokens, and deep mode runs one to four hours by the docs' own estimate, so cost scales with patience. The sandbox image pulls many tools at their latest version at build time, so two builds can differ. Inside the container the agent's user has passwordless sudo, which is reasonable for a pentest box but means the container boundary is the only fence. The command-line entry point imports PostHog and Scarf telemetry modules. The results viewer hands out a tokened link that grants access to the run, and the project itself warns to share it carefully.

How to use it well

Use it as a second opinion on code and apps you own or are authorized to test: quick mode on each pull request, deep mode before a release, as the project's CI guide suggests. Give it the source and the running app together, since the standard and deep modes use source-aware triage to pick what to attack dynamically. Set a spending cap and pick a capable model, because the agents do the reasoning. It does not replace a human-led engagement report on its own, and features like automatic patch pull requests and continuous scanning sit in the hosted service, not the open package.

Observed

License

Apache-2.0

Language

Python, requires 3.12 or newer

Packaging

PyPI package strix-agent with a strix console command; also a curl install script

Declared maturity

Development Status :: 3 - Alpha classifier in pyproject.toml

Runtime requirement

Docker, plus an LLM API key or a local OpenAI-compatible model server

Sandbox

Kali Linux based container with Nmap, Nuclei, SQLMap, ZAP, Semgrep, TruffleHog, Gitleaks, Trivy, Caido and Playwright listed

Model routing

LiteLLM provider/model strings; Bedrock and Vertex as optional extras

Interfaces

Interactive CLI, headless mode, local web viewer, agent skills install, GitHub Actions workflow

CI behavior

Exit code 2 when vulnerabilities are found

Scan modes

quick, standard, deep; deep is the default

Container user privileges

Sandbox user has passwordless sudo inside the container