Warp Agent CLI: Warp's Coding Agent Comes to Any Terminal

> **TL;DR:** Warp has launched the Warp Agent CLI, a standalone command-line coding agent that runs in any terminal — iTerm2, Ghostty, VS Code, Windows Terminal, and more — not just Warp's own app. It ships with multi-model auto-routing, a $18/month plan that includes $20 in monthly inference credit (or pay-as-you-go credits from $10), and the unusual ability to drive interactive terminal apps and even orchestrate rival agents like Claude Code and Codex.

Key Takeaways

- Warp Agent CLI decouples the agent from Warp's terminal: one install command puts it in iTerm2, Ghostty, VS Code, Windows Terminal, or macOS Terminal. - Multi-model auto-routing picks a model per task, with frontier and US-hosted open-weight models included, plus YAML-configurable custom routers and OpenAI-compatible endpoints. - Pricing is $18/month for $20 of monthly inference credit, ad hoc credits from $10 with no subscription, or bring your own API key. - Its terminal DNA shows: it distinguishes shell commands from natural-language prompts and can operate full-screen apps like vim, gdb, htop, and REPLs. - It can delegate to subagents and coordinate other harnesses — including Claude Code and Codex — positioning it as an orchestrator, not just a competitor.

Warp has released the [Warp Agent CLI](https://www.warp.dev/blog/introducing-the-warp-agent-cli-coding-agent), a standalone command-line coding agent that runs in any terminal — not just Warp's own app. The launch takes the agent that already powers Warp's AI terminal and ships it as an installable tool for iTerm2, Ghostty, VS Code, Windows Terminal, and macOS Terminal, putting it in direct competition with Anthropic's Claude Code and the fast-growing field of terminal-native coding agents we cover in [New AI Tools & Skills](https://speka.info/new-ai-tools/).

From Terminal Feature to Standalone Agent

Warp built its reputation on an AI-powered terminal, with its coding agent living inside that app as a headline feature. The Warp Agent CLI decouples the two. A single install command — a `curl` one-liner on Mac and Linux, a PowerShell equivalent on Windows — drops the same agent into whatever terminal you already use.

Under the hood, the CLI inherits real terminal infrastructure rather than a chat loop bolted onto a shell. Warp says it is built on a terminal multiplexing layer similar in architecture to tmux, which enables persistent agent sessions, directory switching mid-session, and remote agent execution without installing a binary on the target machine. The company claims this architecture unlocks "capabilities beyond what is possible in other agents" — a marketing line, but one backed by concrete features covered below.


Multi-Model Routing Is the Headline Feature

Most CLI agents are tied to one vendor's models. Warp's pitch is the opposite: the Agent CLI is a multi-model tool that automatically routes each task to a model based on its complexity, an approach Warp describes as "pareto-efficient" — spend big-model money only where the task warrants it.

Out of the box it includes frontier models alongside US-hosted open-weight models. Power users can go further: custom model routers are configurable via YAML, OpenAI-compatible endpoints are supported, you can bring your own API key, and there is even sign-in support for a SuperGrok subscription. That flexibility contrasts with single-vendor agents, and it lands at an interesting moment — Anthropic has been tuning its own lineup for exactly this kind of workload, as we covered in our look at [Claude Opus 5, Anthropic's model for long-running agents](https://speka.info/blog/claude-opus-5-anthropics-model-for-long-running-agents).

What It Costs

The pricing is straightforward. A Warp subscription runs $18 per month and includes $20 of inference credit monthly — nominally more credit than the subscription costs. If you don't want a subscription, ad hoc credits start at $10. And if you already pay a model provider directly, the bring-your-own-key option sidesteps Warp's metering entirely.

Built Like a Terminal, Not a Chat Window

The features that most distinguish the Warp Agent CLI come from its terminal heritage. The prompt uses natural-language detection to tell shell commands apart from instructions to the agent, so you don't mode-switch between "typing commands" and "talking to the AI." Tab completion works for arguments and flags.

More unusual: the agent can control full-screen interactive terminal applications. Warp lists vim, the sqlite shell, the Python REPL, gdb, and htop as examples. That matters in practice — a large share of real debugging and operations work happens inside interactive TUIs that most coding agents simply cannot see or operate.

An Agent That Orchestrates Other Agents

Perhaps the boldest positioning move: the Warp Agent CLI doesn't just compete with rival agents, it claims to coordinate them. Warp says the tool can delegate work to subagents and orchestrate different harnesses — explicitly naming Claude Code and Codex. There is also a cloud handoff path, letting you transfer an in-progress task to a cloud agent that keeps executing remotely after you close your laptop.

If that orchestration works as advertised, Warp is angling to be the layer above the model-vendor agents rather than a like-for-like alternative — a meaningful strategic difference in a market where the underlying models change quarterly.

Why This Launch Matters

The terminal-agent space is consolidating into a genuine platform race. Model vendors ship first-party agents; now terminal and editor vendors are answering with their own, and Warp — which arguably has the strongest terminal DNA of any of them — just made its play portable. The launch drew substantial developer attention, including an active [discussion on Hacker News](https://news.ycombinator.com/item?id=49171766).

It also arrives amid a broader arms race across the AI stack, from open safety models like [Mistral's Shieldstral moderation model](https://speka.info/blog/mistral-shieldstral-3b-open-model-for-ai-moderation) to the scrutiny of training practices raised by [Anthropic's Project Panama](https://speka.info/blog/anthropics-project-panama-books-destroyed-to-train-claude). For developers, the practical takeaway is simpler: the coding agent you use no longer has to be decided by the terminal you open or the model vendor you pay. Warp is betting that routing, orchestration, and terminal-native control are the features that win that choice.

Frequently Asked Questions

What is the Warp Agent CLI?

It's a standalone command-line coding agent from Warp — the same agent built into the Warp terminal, now installable in any terminal app, including iTerm2, Ghostty, VS Code, Windows Terminal, and macOS Terminal.

How much does the Warp Agent CLI cost?

A Warp subscription is $18/month and includes $20 of inference credit each month. Ad hoc credits start at $10 with no subscription, and a bring-your-own-API-key option is also available.

Which models does the Warp Agent CLI support?

It includes frontier models and US-hosted open-weight models out of the box, with automatic routing based on task complexity. It also supports custom model routers via YAML, OpenAI-compatible endpoints, and SuperGrok subscription sign-in.

How is it different from Claude Code?

Claude Code is tied to Anthropic's models, while Warp's agent is multi-model with automatic routing. Warp also says its CLI can operate full-screen terminal apps like vim and gdb, and can even orchestrate other agent harnesses, including Claude Code and Codex.

Do I need the Warp terminal to use it?

No. The Agent CLI installs via a single command on Mac, Linux, or Windows and runs in whatever terminal you already use.

Sources

- https://www.warp.dev/blog/introducing-the-warp-agent-cli-coding-agent - https://news.ycombinator.com/item?id=49171766