A self-hosted orchestrator that runs a team of AI agents per project: a lead that plans and assigns, workers that build in their own containers, a reviewer and a second model that check each other, and gates that decide when a task is really done. Templates for dev, security, research and marketing work, a real-time HUD, and a phone app so you can steer and approve from any screen.
Open source · 8 team templates · a security team found real bugs in OWASP Juice Shop, a dev team shipped a data pipeline across 6 reviewed PRs
TypeScriptNode.jsReactSQLiteNeo4jDocker
Website ↗ · Case study ↗ · GitHub ↗
Detects malicious Windows PE and Linux ELF files by fusing two models. A static LightGBM model scores 2,568 EMBER features; a behavioral model scores API traces from sandboxed CPU emulation. An out-of-fold logistic stacker combines them. It runs fully offline, with a desktop GUI, a CI-friendly CLI, and folder monitoring.
1,701 labelled PEs · static 0.968 AUC → fused 0.997 AUC · TPR @1% FPR 83% → 98%
LightGBMscikit-learnEMBERSpeakeasyQilingPySide6
A character-level TextCNN that reads raw HTTP requests and flags SQL injection, XSS, path traversal and command injection without signatures. Parallel 3/5/7-width convolutions with global max-pooling make detection position-independent, so payloads are caught in the path, query or body. Served behind a Flask gateway with a live blocking dashboard and an /inspect scoring API.
Trained on CSIC 2010 · class weighting · early stopping on validation AUC
TensorFlowKerasTextCNNFlaskpytest
An LLM agent plans reconnaissance for authorized red-team work, and a human approves each step before it runs. Tool output is parsed into typed records, findings map to MITRE ATT&CK and CVEs, and a scope guard plus an append-only audit log keep it safe. It runs on local or low-cost models (Ollama, NVIDIA NIM, Gemini) and ships with a precision/recall/F1 evaluation harness.
LLM agentsFastAPIOllamaDockerMITRE ATT&CK