One REPL. Two engines under it.

Vanta Agent is one REPL and one frozen contract with two interchangeable engines behind it: Claude through the Agent SDK, running in-process, or OpenAI Codex, driven as a subprocess. Same terminal, same tools, same cards — which engine is underneath is a detail.

Both engines speak the same MCP tools, stream the same tool cards, and report the same way — so the Supervisor drives Claude and Codex identically, and nothing provider-specific leaks through.

Change the engine, model, effort, or auth without restarting the process. /model and /effort carry the same conversation forward; /provider and /auth start a clean thread on the new engine — and tell you so.

Vanta Studio orchestrates the agent exactly like everything else — spawned with bypassed permissions into an isolated worktree, its MCP bearer scoped to its identity. It reports status two ways: a byte-frozen terminal footer and spinner the desktop can scrape, and an opt-in structured NDJSON channel.

The channel is best-effort and additive: with no listener it degrades silently to a no-op, and the terminal contract stays the single source of truth — behavior is byte-for-byte identical to a build without it.

It is a real CLI, not just an internal backend. Install it, point it at any repo, and drive it by hand. Outside bypass mode it asks before it changes anything.

• Diff approval — every edit and command reviewable before it runs (y · n · a).
• @file mentions, slash commands, and your own markdown commands.
• Reads Claude Code's .mcp.json — your vanta-studio tools work out of the box.
• API key or your claude.ai / codex login; pick per launch or persist a default.