A terminal coding agent that refuses to pick a favorite model

What it does

VT Code is an open-source terminal coding agent written in Rust, built on ratatui/crossterm. It offers an interactive TUI with slash commands and streaming responses, plus non-interactive ask and exec modes for scripting. The tooling covers file operations, patching, ripgrep search, syntax-aware code intelligence, and project indexing.

The interesting bit

The provider list is the real flex: 20 providers out of the box, from OpenAI and Anthropic down to Ollama, LM Studio, and llama.cpp, plus any OpenAI-compatible endpoint via [[custom_providers]]. It also runs as both an MCP client and server, and can operate as a subagent under Claude Code or via ACP in Zed — it’s an agent that’s happy being the sidekick.

Key highlights

• Read-restricted shell sandbox: dangerous commands blocked, file ops scoped to the project, all tool calls audit-logged
• Agent Skills using the open SKILL.md format, discovered at repo, user, admin, and system levels
• Protocol support: Open Responses, Agent2Agent, Anthropic Messages API compatibility, ATIF trajectory export
• Ghostty VT snapshots for rich PTY rendering, with a legacy_vt100 fallback
• Session resume/continue and dynamic context curation

Caveats

• The default model is Xiaomi MiMo V2.5 Pro — a sponsorship-flavored default you’ll likely want to change
• Its benchmark entry on the Next.js evals leaderboard is still pending, so no published numbers to judge it by
• Windows builds are explicitly “best-effort” and may lag behind

Verdict

Worth a look if you want a single terminal agent that works with whatever model you already pay for, local or cloud. If you’re settled on Claude Code or Codex CLI, the pitch is mostly “same idea, more providers” — the pending benchmarks don’t yet argue otherwise.