Warp's Universal Agent Support: The Terminal as an Agentic Development Environment

Warp's universal agent support wires third-party CLI coding agents into the terminal as first-class citizens rather than generic shell processes. It launched on April 14, 2026 with Claude Code, Codex, Gemini CLI, and OpenCode; Warp's docs now list 16 supported agents. Together with vertical tabs, Warp's pane-stacking surface for running several agents side by side, it makes Warp's pitch concrete: the terminal is a dashboard for several AI coding agents working in parallel. Warp's name for the category is Agentic Development Environment, or ADE, and since April 2026 the client itself is open source under AGPL v3.

Key Takeaways

• Universal agent support covers 16 CLI agents, including Claude Code, Codex, OpenCode, Gemini CLI, Copilot CLI, Cursor CLI and Antigravity CLI
• Vertical tabs show each pane's agent, branch and status; only Claude Code, Codex and OpenCode get agent notifications
• Code review pane sends inline comments straight back into the running agent without copy-paste
• Remote Control publishes a running agent session so you can monitor and steer it from your phone or another computer
• Third-party agents bill to their own accounts; Warp's plans pay for Warp's own agent and cloud runs
• Pricing: Free; Build $20/mo, Max $200/mo, Business $50/user/mo (10% less billed annually)
• The Warp client is open source under AGPL v3

What Is an Agentic Development Environment?

The Agentic Development Environment label is Warp's framing for what a developer surface looks like once AI coding agents stop being a side panel and start being the main thing you do. Zach Lloyd, Warp's CEO, stated the positioning bluntly at the Warp 2.0 launch in June 2025: "Not a terminal, not an IDE. Warp is the Agentic Development Environment, built for coding with multiple AI agents." His argument is that IDEs bolt agents onto code editors through chat panels, while CLI tools bury them inside a single shell stream, and neither is native to the agentic workflow.

The April 2026 launch is the second proof point of that thesis. Warp 2.0 introduced the framing; universal agent support is what makes the framing tangible. You sit down, open Warp, and the surface assumes from the first keystroke that you are about to start several agents at once, on several branches, talking to several models. The terminal is the outer chrome. The agents are the work. The terminal-as-ADE move is one of the defining shifts in the 2026 AI tools landscape.

First-Class CLI Agent Support: What "First-Class" Adds

Warp's third-party CLI agents docs list 16 supported agents: Claude Code, OpenAI Codex, OpenCode, Amp, Auggie, Copilot CLI, Cursor CLI, Gemini CLI, Droid, Pi, Goose, Antigravity CLI (agy), Hermes, Mistral Vibe, Grok Build, and Kiro CLI. Warp detects a supported agent when you launch it in a pane and turns on its agent toolbelt. Any of them will run inside any terminal that can spawn a subprocess, so what does "first-class" actually buy you?

The difference is everything Warp wraps around the agent process. Every supported agent gets the rich input editor (Ctrl-G): multiline prompts, mouse selection, cut, copy, paste, voice input, image attachment, a file explorer, and saved prompts and skills, instead of the reduced affordances of a raw subprocess. Every one gets code review comments, attach-code-as-context, vertical tabs with metadata, Tab Configs, and Remote Control. Agent notifications (complete, needs input, error) are the exception: Warp's feature table lists them only for Claude Code, Codex, and OpenCode, each after a one-time plugin or config setup.

One caveat for Gemini CLI users: Google stopped serving Gemini CLI to free, Google AI Pro and AI Ultra users on June 18, 2026 and moved them to Antigravity CLI, which Warp also recognizes. Our guide to Google's AI stack covers who keeps Gemini CLI.

Run the same claude or codex binary in plain Terminal.app and you get the agent. Run it in Warp and you get the agent plus the dashboard around it. The point of the launch is that the dashboard is now the unit of work.

Vertical Tabs and the Orchestrator Dashboard

Vertical tabs stack each pane as a row in a sidebar list, and each row carries enough metadata to act as a status board. The vertical tabs documentation lays out the visual contract precisely.

Rows are not columnar. In compact mode, the default, each pane row shows an icon and a title on one line, with an optional subtitle below it. The title is configurable: the last command or conversation (the default), the working directory, or the git branch. The subtitle, set through the "Additional metadata" option, shows one of branch, working directory, or command/conversation, whichever the title is not already using. Expanded mode gives each row a title, a description such as the working directory, and metadata: the git branch, a diff-stats badge, and a PR badge when available. The PR link and diff stats are expanded-mode only, and the PR link requires the GitHub CLI. Double-click a row to rename a tab inline. Panes inside a tab cannot yet be renamed individually.

The status badges are the thing. Each row shows a colored icon that tells you what the agent in that pane is doing without you having to switch into it:

• in-progress (magenta clock): the agent is actively thinking or running a tool
• done (green check): the last turn finished cleanly
• error (red triangle): the agent or its tool returned an error
• blocked / awaiting approval (yellow stop): the agent needs you to approve a permission request before it can continue
• cancelled (gray stop): you stopped the run

An accent dot on the row marks unread agent activity since you last looked at the pane. The combined effect is a glanceable dashboard: ten panes, ten status dots, your eye lands on the yellow stop and you know exactly which agent needs your attention next. Warp's vertical tabs announcement frames this as the difference between juggling tabs and running an operations center.

Tab Configs sit on top of the surface as one-click templates. A Tab Config can pre-load a working directory, a set of startup commands, a theme, and a worktree, so spinning up a new pane on a new branch is a single click rather than three commands.

Code Review Without Copy-Paste

Reviewing an agent's work has historically meant flipping between the terminal where the agent is running and a diff viewer somewhere else, then copy-pasting feedback back into the prompt. Warp's code review integration removes the round-trip.

The review pane lets you send inline review comments, file selections, and snippets straight back into the running CLI agent session without leaving Warp. You read the diff, highlight the function you want changed, type your comment, and send; the comment lands in the agent's input stream as if you had typed it into the prompt. The agent sees the file context, the selection, and your note in one unit, and replies in the same pane.

The practical effect is that the inner loop of "agent proposes, you review, agent revises" stays inside one window. You stop losing context to tab-switching, and the agent's view of your feedback stays attached to the code you were looking at when you wrote it.

Remote Control: Steer Agents From Your Phone

Warp's Remote Control publishes a running third-party agent session, such as Claude Code, Codex, or OpenCode, to the cloud with one click, per its docs. Click the /remote-control chip in the agent utility bar and Warp copies a shareable link; from a phone, a browser, or another computer you can then send input, approve commands, or redirect the agent. Teammates you share the link with can watch or steer too. Syncing stops when you close or stop publishing the session.

The flow is built for the realistic shape of long-running agent work: you start a Claude Code or Codex session before you stand up from your desk, the agent grinds for 20 minutes, hits a permission prompt, and waits. Without Remote Control, the agent sits idle until you are back at the laptop. With it, your phone is enough. You see the prompt, approve or reject, and the agent keeps moving.

This is the same operational shift that scheduled agents introduced from a different angle. For unattended runs that fire on a cron or a webhook, see Claude Code Routines and the comparison of scheduled-agent platforms for how each tool handles the "agent works while you are away" problem.

A Day in the Life With Warp

Picture a feature day with universal agent support and vertical tabs in front of you. You open Warp and trigger four Tab Configs in sequence, each pre-wired to a different worktree on a different branch. One pane runs Claude Code on the API refactor, one runs Codex on the test suite, one runs Copilot CLI on the docs pass, one runs OpenCode on a small migration script. Four vertical tabs, four status dots, four branches.

You glance at the sidebar. The API refactor pane shows a magenta clock: still thinking. The test suite pane shows a green check: done, ready for review. The docs pane shows a yellow stop: waiting on your approval to write a file. You click the docs pane, approve the write, and click straight to the test suite pane to read the diff. You highlight a function in the review pane, type a comment about a missing assertion, and send it back into Codex without leaving the window. The migration script pane finishes while you are looking elsewhere, the accent dot lights up to tell you, and you queue the next prompt.

You walk to lunch after publishing the API refactor session with Remote Control. On the way, it hits a permission request; you open the link on your phone, approve from the sidewalk, and the agent keeps grinding. The shift this makes concrete is from "one agent in one terminal" to "several agents on one dashboard." The mental model will be familiar to anyone who has juggled multiple Claude Code windows; the comparison between Cursor and Claude Code walks through how the same orchestrator role plays out across editors.

Open Source Under AGPL v3

Warp's client is open source. The warpdotdev/warp repository is licensed under AGPL v3, except the warpui_core and warpui UI framework crates, which are MIT. Warp open-sourced the client in April 2026, alongside a new settings.toml file; the server side stays closed. For agent users, that means the process model, telemetry and agent integration that wrap your Claude Code or Codex session are inspectable. Our coverage of the open-source release covers the license split, building from source, and what the TOML settings file changes.

Pricing and How Usage Is Billed

Warp's pricing page lists four self-serve plans and states each allowance in dollars of agent usage at API rates:

Enterprise is custom. Warp's billing docs still meter that usage as credits in three buckets: AI credits for the model call when Warp pays for it, compute credits for Warp-hosted cloud sandboxes, and platform credits for cloud runs and for local runs on Business and Enterprise that use your own inference.

The important point for universal agent users is what is not billed. Warp's docs say running claude, codex, or another agent CLI yourself, "whether in a Warp terminal pane or elsewhere, doesn't consume platform credits," and AI credits apply only when Warp pays for the model call. Claude Code bills to your Anthropic plan or API key, and Codex to your ChatGPT plan or OpenAI key, exactly as they would in any other terminal. Warp's docs attach no plan requirement to first-class agent support: run a supported agent and the toolbelt appears. What a paid Warp plan buys is usage for Warp's own agent, the Warp Agent, and its cloud agents, so a team that only runs third-party CLIs can stay on Free and pay its agent vendors directly.

Corrections and Updates

• October 9, 2026: This page listed Build at $18/mo with 1,500 credits, Max at $180 with 12x credits and Business at $45, and said Warp's credits back the LLM calls each agent makes; those were annual-billing prices, monthly prices are $20, $200 and $50 with usage stated in dollars, and third-party CLI agents bill to their own accounts, not Warp's.

Sources

• Warp blog, "Introducing Universal Agent Support: level up any coding agent with Warp", Zach Lloyd, April 14, 2026. The launch date, the four launch agents, and the feature set. Checked October 9, 2026.
• Warp docs, "Third-party CLI agents overview". The 16 supported agents and the per-agent feature table. Checked October 9, 2026.
• Warp docs, "Remote Control". Publishing and steering a session from another device. Checked October 9, 2026.
• Warp pricing. Free, Build, Max, Business and Enterprise plans, monthly and annual prices, and included agent usage. Checked October 9, 2026.
• Warp docs, "Warp credits and billing" and "Platform credits". The three credit buckets and the rule that third-party agent CLIs do not consume Warp credits. Checked October 9, 2026.
• warpdotdev/warp on GitHub. AGPL v3 and MIT license split. Checked October 9, 2026.
• Warp blog, "Introducing Warp 2.0: Reimagining coding with the Agentic Development Environment", Zach Lloyd, June 24, 2025. The ADE framing and Zach Lloyd's quote.
• Warp docs, "Vertical tabs". Row layout, metadata and status badges.
• Google Developers Blog, "Transitioning Gemini CLI to Antigravity CLI", May 19, 2026. The June 18, 2026 Gemini CLI cutoff for individual users. Checked October 9, 2026.