Strix runs autonomous penetration-testing agents against source code, web applications, APIs, and infrastructure. The agents execute the target dynamically rather than reading it statically, and each finding is validated with a working proof-of-concept exploit and reproduction steps. It is aimed at developers and security teams who want testing without a manual engagement, and it reports findings with remediation guidance.
Runs start from the command line against a directory, a repository URL, a live application, or an API contract such as an OpenAPI, Swagger, or Postman collection paired with a base URL; several targets can be combined so source code and a deployed application are tested together. The first run pulls a sandbox container image, and results are written to a per-run directory. A non-interactive mode prints findings and the final report and exits with a non-zero code when vulnerabilities are found, which is how the supplied GitHub Actions workflow gates a pull request; in a pull-request run, quick reviews are scoped to changed files. A local viewer serves the findings, a live map of the agent team, and past runs from disk, bound to localhost behind a tokened link.
The agents carry an offensive-security toolkit: an HTTP interception proxy, an automated browser for client-side testing, an interactive shell, a Python sandbox for writing exploits, reconnaissance and OSINT tooling, static and dynamic analysis, and a structured findings store with CVSS scoring and OWASP classification. Multiple specialised agents run in parallel for reconnaissance, exploitation, and post-exploitation, and share discoveries so vulnerabilities can be chained.
A model is selected with the STRIXLLM and LLMAPIKEY environment variables, and configuration is saved to a local file between runs. Documented options include a ChatGPT subscription sign-in instead of a metered key, a local model endpoint through LLMAPI_BASE, and Vertex AI, Bedrock, and Azure among the supported providers. Model Context Protocol servers listed in a local JSON file are exposed to the agent during a run, with their tools namespaced and filterable. A managed cloud runs scans without Docker or a model key, driven either from the web application or with strix cloud subcommands that mirror the REST API.
Features
- Validated findings: every reported vulnerability comes with a working proof-of-concept and reproduction steps
- Multi-agent runs: specialised agents for reconnaissance, exploitation, and post-exploitation run in parallel and share discoveries
- Pentesting toolkit: HTTP interception proxy, browser exploitation, shell access, a Python exploit sandbox, and OSINT tooling
- Coverage: access control, injection, server-side, client-side, business-logic, authentication, infrastructure, and API classes
- API contract testing: point it at an OpenAPI, Swagger, or Postman source plus a live base URL to test declared endpoints
- CI integration: headless mode exits non-zero on findings, with a GitHub Actions workflow that scopes quick reviews to changed files
- Local viewer: strix view serves findings, an agent map, and past runs from disk on localhost behind a tokened link
- Coding-agent skills: nine installable skills let a skill-compatible agent run pentests, fix findings, and set up CI scanning
- MCP client: local stdio or remote HTTP MCP servers can be listed and their tools exposed to the agent, with filtering
- Managed cloud: strix cloud subcommands mirror the REST API for scans, vulnerabilities, and source uploads
