VELA

VELA is a secure code execution runtime designed specifically for AI agents and SaaS platforms. It allows developers to run untrusted, AI-generated code safely by isolating each execution in a Firecracker micro-VM. This provides hardware-level security, ensuring that if the AI model is compromised or generates malicious code, it cannot affect your host system. Each code run is hermetically sealed and starts in about 150 milliseconds from a pre-warmed pool of VMs. The system uses HMAC-SHA256 signed capability tokens that define exactly what each caller can do—specifying allowed filesystem paths, network access, memory limits, and timeouts. It also includes a policy engine where you can define YAML deny lists to block specific operations like 'rm -rf' or 'curl', and can require human approval for sensitive actions. VELA stands out by combining enterprise-grade security with developer-friendly integrations. It offers ready-to-use adapters for popular AI frameworks like LangChain, LlamaIndex, CrewAI, and OpenAI function tools, making it simple to drop into existing AI agent stacks. Every execution creates a full audit trail written to an append-only JSONL log, which is also broadcast to a live stream socket and queryable by ID. This tool is most valuable for developers building AI agents that need to execute code, ML platform engineers responsible for production safety, and SaaS companies offering code execution features. Real-world use cases include safely running code generated by language models, replacing complex Docker orchestration with a simpler, more secure alternative, and adding a protected code interpreter to AI assistant products.