The Secure AI Agent Framework — build production autonomous agents in Rust with sandboxed skills, 6-layer guardrails, and 8 LLM backends out of the box.
Each capability ships as a focused crate — include only what your agent needs.
6-layer pipeline: PII detection, prompt injection, toxicity, shell injection, base64 exfiltration, and unicode smuggling — all before any LLM call.
Real HTTP multimodal calls to Claude, OpenAI, and Gemini. Pass images, audio, and documents. Streaming responses out of the box.
Built-in web fetch, search, and extraction. Computer use skill for screenshot, mouse, and keyboard control. No external dependencies.
Orchestrator-Workers pattern per Anthropic guidance. TaskQueue, AgentMonitor, agent profiles, and human-in-the-loop approval flows.
First-class Python bindings via PyO3. Run Argentor agents from Python scripts, notebooks, or FastAPI services — full feature parity.
MCP client (JSON-RPC 2.0 stdio) ships as a core skill. WASM sandboxed plugins via Wasmtime — untrusted code runs isolated, always.
Run the hello-world example in under a minute — no API keys required for the demo backend.
Measured against a Python/LangChain baseline on equivalent pipelines. Full methodology →
Results from benchmarks/ — run cargo bench to reproduce on your hardware.