Show HN: VT Code – LLM-agnostic coding agent with MCP/ACP and sandboxed tools (github.com)

VT Code is a new Rust-based terminal coding agent released as a CLI and prebuilt binaries (cargo/brew/npm/GitHub Releases) that combines semantic code intelligence (Tree-sitter parsers for Rust, Python, JS/TS, Go, Java) with ast-grep structural search/transform and multi-LLM orchestration. It natively supports OpenAI, Anthropic, Gemini, xAI, DeepSeek, Z.AI, Moonshot, OpenRouter and local Ollama models, with automatic failover, prompt caching, and provider-agnostic abstractions to keep model IDs and configs reproducible. Safety and usability features include sandboxed execution, path validation, per-tool allowlists/prompts, and Zed IDE integration via the Agent Client Protocol (ACP). Under the hood VT Code uses Tokio for multi-threaded async execution, tiktoken-rs for accurate token budgeting, and a modular tool/LLM trait system enabling streaming responses and custom tools. Its context engineering system detects conversation phases, performs iterative curation, triggers LLM-driven summarization (target ~30% compression; reported ~29% token savings/turn), and enforces thresholds (warn at 75%/compact at 85%) to support long-running sessions with minimal token overhead. Additional technical highlights: MCP support for domain tool discovery, decision ledgers for auditability, AST-aware refactors with preview mode, and extensible security policies. For AI/ML practitioners and agent researchers, VT Code is a performant, reproducible platform to experiment with semantic code agents, multi-provider orchestration, and real-world safety/efficiency patterns.