kritt-ai/open-kritt

open-kritt is a self-hosted platform that puts AI agents to work finding security vulnerabilities in code. Instead of pointing one AI model at an entire codebase and hoping it spots problems, which the project says rarely works well, it breaks security research into small, focused tasks and runs many AI agents on those tasks in parallel. The results are then combined, checked for duplicates, and ranked by severity so a human researcher can review a manageable list of findings rather than raw noise. The tool is aimed at security researchers and developers who care about security, particularly people who want control over their own prompts, workflows, and choice of AI provider. You can build reusable workflows that chain prompts together into research playbooks, run scans against remote or local code repositories using Codex or Claude Code, and use post-scan scripts to verify that a finding is real, build a proof of concept, and generate a report. It supports connecting to OpenAI, Anthropic, or OpenRouter for model access, or using a Codex login. The people behind open-kritt say they have earned over 1.5 million dollars in bug bounty payouts under a researcher identity, and describe this project as the open version of the internal tooling behind that work. Setup requires Git, Docker with Docker Compose, and Node.js version 20 or newer. After cloning the repository, running a setup script and a start script brings up the full application, which you access through a local web browser at port 5173. The README is explicit that the backend has no built in authentication and should only run on a private, dedicated machine, since the AI agents run with root access and direct internet access inside disposable containers so they can install tools and compile code during a scan. The project is licensed under the GNU Affero General Public License version 3.