VELA.

Vela

Vela is a secure execution runtime that protects AI agents and SaaS platforms from untrusted code. Built on Firecracker micro-VMs with HMAC capability tokens and full audit trails, it delivers hardware-level isolation without container overhead—giving you complete control over what AI-generated code can do.

Product Highlights

• Firecracker Micro-VM Isolation: Hardware-level VM isolation per execution with ~150ms p50 cold starts from pre-warmed pools—nothing leaks between runs and your host OS remains untouched.
• HMAC Capability Tokens: Cryptographically signed tokens that declare exact permissions including filesystem paths, network access, memory limits, and timeouts—eliminating guesswork about what code can execute.
• Full Audit Trail: Every execution event written to append-only JSONL logs with live stream broadcasting and ID-based querying—complete observability for compliance and debugging.
• Framework Adapters: Native integrations for LangChain, LlamaIndex, CrewAI, and OpenAI—drop-in replacements that secure your existing AI stack without rewrites.
• Policy Engine: YAML-based deny lists to block dangerous patterns like rm -rf or curl, with optional human approval gates for sensitive operations.
• Real-Time Intrusion Detection: Automatic flagging of syscall spikes, sensitive file access, and private network connections with structured alerts.

Use Cases

• AI Agent Code Execution: Safely run LLM-generated Python or JavaScript without exposing your infrastructure to prompt injection or jailbreak attacks.
• Multi-Tenant SaaS Platforms: Isolate customer code execution with per-request micro-VMs, eliminating cross-tenant data leakage risks.
• Secure CI/CD Pipelines: Execute untrusted build scripts and test code with granular capability controls and complete execution logs.
• Fintech & Healthcare Compliance: Meet strict audit requirements with immutable execution records and hardware-isolated environments for sensitive computations.

Target Audience

Vela is designed for engineering teams building AI-powered applications, ML platform engineers, and security-conscious developers who need to execute untrusted code without compromising infrastructure security or observability. It serves organizations from early-stage startups to enterprises that cannot risk raw subprocess execution or tolerate the overhead and opacity of hosted sandbox solutions.