VT Code is what happens when a coding agent takes security seriously
Yet another terminal coding agent? Yes. But VT Code is the first one that treats sandboxing as a first-class feature instead of an afterthought, and that matters more than you'd think after watching this month's Claude Code regression debate.
The project showed up on HN today at 522 stars, written in Rust, runs as a TUI. It supports OpenAI, Anthropic, DeepSeek, Gemini, Ollama out of the box. The differentiation is the trust model. Bash commands get parsed by tree-sitter and run through execution policies before they touch your shell. The agent itself runs inside macOS Seatbelt or Linux Landlock+seccomp by default. Any destructive action requires human approval. None of this is opt-in.
More importantly, VT Code is one of the first independent agents to ship the new Agent Client Protocol (Zed integration) and Agent2Agent protocol (so other agents can talk to it). It also implements the Agent Skills standard. That's three open protocols where a typical Cursor-clone implements zero. The bet is that the agent ecosystem fragments into specialists and orchestrators, and the agents that survive are the ones that speak the lingua franca.
If you've been waiting for a Claude Code alternative that won't quietly rm -rf your repo on a hallucinated bash command, this is the one to try this weekend.
https://github.com/vinhnx/vtcode
← Back to all articles
The project showed up on HN today at 522 stars, written in Rust, runs as a TUI. It supports OpenAI, Anthropic, DeepSeek, Gemini, Ollama out of the box. The differentiation is the trust model. Bash commands get parsed by tree-sitter and run through execution policies before they touch your shell. The agent itself runs inside macOS Seatbelt or Linux Landlock+seccomp by default. Any destructive action requires human approval. None of this is opt-in.
More importantly, VT Code is one of the first independent agents to ship the new Agent Client Protocol (Zed integration) and Agent2Agent protocol (so other agents can talk to it). It also implements the Agent Skills standard. That's three open protocols where a typical Cursor-clone implements zero. The bet is that the agent ecosystem fragments into specialists and orchestrators, and the agents that survive are the ones that speak the lingua franca.
If you've been waiting for a Claude Code alternative that won't quietly rm -rf your repo on a hallucinated bash command, this is the one to try this weekend.
https://github.com/vinhnx/vtcode
Comments