Preface¶

AI programming assistants in terminals can be roughly divided into two categories in the past: one is Copilot and Cursor integrated into IDEs; the other is “command-line Agents” such as Claude Code and Codex CLI. Previously, Warp integrated Agents into its own terminal, Warp Terminal, and users had to use Warp to access that multi-model orchestration capability. On August 4, 2026, Warp’s official blog announced the launch of Warp Agent CLI — spinning off the programming Agent from Warp Terminal into a CLI that can run on any terminal, available on Ghostty, iTerm 2, VS Code built-in terminal, and Windows/macOS system terminals. This article verifies key points based on Warp’s official announcement and CLI documentation, sorting out its capability boundaries and impact on the terminal Agent track.

From Warp Terminal to Standalone CLI¶

Warp Agent CLI shares the same Warp Agent inference framework with the Warp app: account, subscription, model permissions, Rules, and Skills are consistent across both ends; installing the Warp terminal app is not mandatory. The startup command is warp, and sessions are rendered as scrolling transcripts in ordinary terminals, supporting streaming output, code diff, tool calls, and inline permission confirmation.

For “heavy terminal users”, the significance of the standalone CLI lies in: no need to switch terminals for the Agent. The official explicitly named common emulators such as Ghostty and iTerm 2, and emphasized that the CLI is a regular command-line program, and SSH remote usage is also supported — this contrasts with the path where “Agents can only be tied to specific IDE/terminal products”.

Core Capabilities: Multi-Model Routing and Cost Optimization¶

Warp positions the Agent CLI as a multi-model, cost-optimizing harness for professional developers. Out-of-the-box capabilities include:

1. Automatic model routing: Automatically selects between frontier models and US-hosted open-weight models based on task complexity, avoiding using the most expensive model for simple tasks.
2. Custom Router: You can define “which model to use for which type of task” through a configuration file, and the documentation provides a Custom router configuration example.
3. Multiple inference access methods (verified by official blog):
   - Warp subscription: Approximately \(18/month**, including approximately **\)20 of inference credits; existing Warp subscriptions can be logged in directly for use.
   - Pay-as-you-go Credits: Starting at approximately $10, no monthly subscription required.
   - Bring your own API Key, OpenAI-compatible endpoints, or SuperGrok subscription login.

The slash command /model can switch the base Agent model; /cost can view the credit consumption of a single response. For individual developers and small teams who care about inference bills, “routing + open-weight models + custom Router” is the differentiator repeatedly emphasized in this release.

Terminal Native: Mux Architecture and Deep Shell Integration¶

Warp claims that other Agent CLIs have gaps in depth of integration with the Shell itself; its solution is to build the CLI Agent on top of Warp’s terminal infrastructure, acting as a built-in mux (multiplexer) for Agent sessions, with an architecture similar to tmux: there is a PTY layer for indirect management between the Agent and the underlying Shell.

The unlocked capabilities (all have official demos or documentation explanations):

Persistent Sessions and Remote Agents
- Switch the working directory in an Agent session without interrupting the session.
- No need to install a remote Agent binary on a remote machine to use it in SSH scenarios — suitable for scenarios such as multiple repositories and restricted cloud host permissions.

Full-screen and Interactive Programs
- Run full-screen/interactive programs such as sqlite, Python REPL, gdb, htop, vim in the same session, and drive or assist them with the Agent (for example, generate SQL in the REPL, set breakpoints in gdb).
- The conversation transcript is retained after the program exits.

Input and Completion
- Agent-first by default: Press Enter to send natural language; enter ! at the beginning of the line to enter Shell mode (the prompt changes from > to !), and return to Agent mode after executing the command.
- Built-in natural language detection classifier: Automatically distinguishes “input that looks like a Shell command” from “Agent prompts” to reduce false triggers.
- Inherits Warp Terminal’s Tab completion: Command, parameter, and path completion are available in Shell mode.

For back-end, SRE, and data engineers who are used to “terminal as the main battlefield”, these capabilities are closer to real workflows than a simple “chat box that can write code”.

Cloud Agent and Multi-Agent Orchestration¶

The CLI connects to the cloud through Warp’s Oz platform: local conversations can be handoff to Cloud Agent, and tasks will continue in the cloud after closing the laptop; you can also monitor and steer on the Oz Web end of the browser, or pull back history and cloud runs using /conversations or warp --resume on the CLI.

Multi-Agent Orchestration is another key point:
- The orchestrator Agent can split complex tasks to sub-Agents for parallel execution; use the arrow keys in the CLI to switch between the orchestrator and sub-Agent sessions.
- Cloud sub-Agents can use a different harness than the parent Agent, and official examples include Claude Code and Codex — that is, Warp’s orchestration layer can schedule heterogeneous Agent runtimes, not limited to its own model stack.
- Local sub-Agents always run on the user’s machine as independent sessions, and use Warp Agent by default.

In terms of slash commands, the documentation lists /agent, /plan, /orchestrate, etc.; MCP servers can be managed via /mcp. This is consistent with the 2026 product narrative of “Cloud Agent + Multi-Agent Factory”, and Warp has sunk the same set of capabilities from the graphical terminal to a general CLI.

Installation and Quick Start¶

macOS / Linux (official one-click script):

curl -fsSL https://app.warp.dev/download/agent-cli | bash

Windows PowerShell:

Invoke-RestMethod "https://app.warp.dev/download/agent-cli.ps1" | Invoke-Expression

Running warp for the first time will go through device authorization login (browser link + verification code); for non-interactive environments such as CI, you can use WARP_API_KEY or --api-key. Common operations:

# Start an interactive session in the current directory
warp

# Resume a specific conversation
warp --resume <conversation-id>

In the session: /version to check the version, /export-md to export Markdown, /compact to compress the context, /auto-approve to toggle whether to automatically approve tool calls. See CLI Reference for complete flags and shortcut keys.

Track Competition: “Infrastructureization” of Terminal Agents¶

The terminal programming Agent track was already crowded in 2026: Claude Code, OpenAI Codex CLI, Gemini CLI, Cursor CLI and others are competing for the “default developer entry point”. Warp’s entry point is terminal infrastructure + multi-model routing + Oz cloud orchestration, rather than making another pure chat-based code assistant:

Whether it is “superior” needs to be verified by actual workloads; from official materials, what can be confirmed is: Warp has productized the Agent capabilities originally locked in Warp Terminal into a first-class CLI citizen, and tied it together with the Oz cloud factory and multi-harness orchestration. For teams already using Warp subscriptions or valuing terminal-native experience, it is worth installing the official script for comparative evaluation; for users still tied to a single vendor’s model, API Keys and OpenAI-compatible endpoints also lower the trial threshold.

Summary¶

Warp Agent CLI was officially released on August 4, 2026, spinning off the multi-model programming Agent from Warp Terminal into a standalone warp command, supporting Ghostty, iTerm 2, VS Code Terminal and the system’s built-in terminal. Core selling points include: task-level model routing, US-hosted open-weight models, deep Shell integration under PTY mux, and Cloud Agent handoff and multi-Agent orchestration (including Claude Code/Codex harness) implemented via Oz. Installation can be done with a single curl command; inference can be accessed via Warp subscription, pay-as-you-go Credits, or your own API Key. The competition for terminal Agents is shifting from “who can write code” to “who understands terminal workflows and multi-Agent scheduling better” — this release by Warp has brought the latter battle to the forefront.