Introducing the Warp Agent CLI: a CLI coding agent that does what others can't
Warp
Warp just shipped its terminal AI agent as a standalone CLI you can run in any terminal, not just their own app. It taps into Warp's tmux-like infrastructure, so it can drive full-screen apps, switch directories mid-session, and hop across SSH boxes with zero remote install.
Warp built its reputation as a terminal app with a built-in AI agent, but starting today that agent doesn't need Warp Terminal at all. The company has pulled it out into a standalone CLI that runs in Ghostty, iTerm 2, VS Code's integrated terminal, or the stock terminal on Windows and Mac. If you spend your day in a shell rather than an IDE, this is aimed squarely at you.
What separates it from Claude Code or Codex isn't the model selection, though it does auto-route between frontier and open-weight models based on task difficulty. It's the plumbing underneath. Warp built the CLI on the same terminal infrastructure that powers its main app, which means the agent manages pty connections with a layer of indirection between itself and the shell, similar in spirit to how tmux multiplexes sessions. Practically, that lets you switch directories mid-conversation, keep a session alive on a remote machine without installing any binary there, and let the agent actually drive interactive programs like sqlite, gdb, htop, or even vim, rather than just shelling out one-off commands.
There's a natural-language classifier baked in too, so typing a plain English request and typing an actual shell command both get routed correctly without you needing to prefix anything, aside from the option to force shell mode with an exclamation mark. Warp also carried over its tab-completion feature from the terminal app, which is a small thing but genuinely useful when you're juggling flags across a dozen CLI tools.
The bigger swing is on orchestration. The CLI can spin up subagents to break a task into pieces, and you can arrow-key between the orchestrator and its subagents to see what each one is doing in real time. Pair it with Warp's cloud platform and you can delegate across not just different models but different agent harnesses entirely, including Claude Code and Codex running as subagents. Start something locally, then hand it off to the cloud and close your laptop; everything gets tracked centrally and you can steer it from a browser later.
Pricing starts at $18 a month for $20 of inference credit, there's a pay-as-you-go option from $10 with no subscription required, and you can also bring your own API key, point it at an OpenAI-compatible endpoint, or log in with a SuperGrok subscription instead.
My take
The multiplexing trick is the real story here, not the multi-model routing everyone else already does. Letting an agent drive vim or gdb interactively, rather than just issuing discrete commands, is the kind of unglamorous infrastructure work that actually changes daily workflow instead of just winning a benchmark. Every other CLI agent vendor is going to be scrambling to copy this by next quarter, and rightly so.
Read more about this at: Warp