What is the 'CLI-everything' wave?

It's the 2026 pattern where AI coding tooling moved out of GUI-first IDEs and into the terminal. The wave has two parts: first-party AI coding CLIs (Claude Code, Codex CLI, Gemini CLI, Aider, Cline, Warp, opencode, Goose, Qwen Code, Cursor Agent, and 20+ more), and second-order platform CLIs (Stripe, Supabase, Vercel, PostHog, ElevenLabs, Resend) that rebuilt their CLIs to be agent-aware so the AI tools above can use them as deterministic action surfaces.

Why are AI tools moving to the CLI instead of staying in IDEs?

Three reasons that matter. First, CLIs compose — pipes, scripts, and exit codes are deterministic ways for one agent to call another. Second, CLIs run anywhere — remote SSH, CI, Docker, ephemeral sandboxes — without packaging an Electron app. Third, CLI surfaces give agents a stable text-only interface where every token of input and output is auditable, which the IDE's rich event stream doesn't offer.

How many AI coding CLIs are there in 2026?

The most comprehensive published comparison maps 30+ AI coding CLIs in 2026. The five most-discussed are Claude Code (Anthropic), Codex CLI (OpenAI), Gemini CLI (Google), Aider (open-source), and Cline (formerly Claude Dev — now Q, the company). Warp 2.0 is in a category of its own as the first 'agentic terminal' rather than a CLI tool you run inside a terminal.

Why did Stripe and Vercel update their CLIs in 2026?

The platform CLIs (Stripe, Supabase, Vercel, PostHog, ElevenLabs, Ramp, Resend, Valyu) updated their CLIs in 2026 to make them more agent-readable: structured JSON output, predictable error codes, idempotent operations, scoped tokens. The agent-to-CLI handoff is now a first-class integration surface, not an afterthought. This is the 'second-order wave' — AI coding agents need deterministic tools, so the tools themselves got upgraded.

Which AI coding CLI should I install first?

If you're on Anthropic, install Claude Code. If you're on OpenAI, install Codex CLI. If you want vendor-agnostic, install Aider or opencode (75+ providers). If your work is mostly shell + scripts, evaluate Warp 2.0 as a full terminal replacement. The honest answer is that the differences between the top 4-5 CLIs are smaller than the differences between any of them and an IDE-first workflow. Pick one, ship for a week, then decide.

Does RuleSell work with all 30+ AI coding CLIs?

RuleSell's catalog is designed around the cross-tool Markdown-rules pattern that most of these CLIs read. A well-written ruleset typically works in Claude Code, Cursor, Cline, Aider, Codex CLI, Gemini CLI, Warp, Windsurf, Continue, and other tools that accept Markdown agent config — with minor format adaptation per tool. The catalog explicitly markets to multi-CLI workflows, and creators publish once for the whole stack.

Topic · A9

The CLI-Everything Wave: 30+ AI Coding CLIs and How to Choose

In 2026, AI coding moved out of IDEs and into the terminal. There are now 30+ AI coding CLIs (Claude Code, Codex, Gemini CLI, Aider, Cline, Warp, opencode, Goose, Qwen Code, Cursor Agent, and more), and the platform CLIs from Stripe, Supabase, Vercel, PostHog, and Resend all shipped agent-aware updates in response. Here's the map, the why, and how to pick.

In 2026, AI coding moved out of IDEs and into the terminal.

This is the meta-trend that's bigger than any single tool. Six months ago, the conversation was "Cursor vs Copilot." Today, it's "which of the 30+ AI coding CLIs do you run, and which platform CLIs have you upgraded to match." The shift happened fast enough that even people inside the industry are still mapping it. This page is the map.

Part 1: the first-party AI coding CLIs

The DEV community published a comparison of 30+ AI coding CLIs in April 2026. The top five by mindshare:

Claude Code (Anthropic) — the most-discussed, ships skills/hooks/MCP as first-class concepts. Best when you're already on Anthropic.
Codex CLI (OpenAI) — OpenAI's official CLI, paired with the GPT-5 series. Best when you're already on OpenAI. Half of the "dual-wield" stack with Claude Code (see /topic/dual-wielding-codex-claude-code).
Gemini CLI (Google) — Gemini-only, reads GEMINI.md as its primary config file. Best for Google Cloud-aligned teams.
Aider (open-source, paul-gauthier) — vendor-agnostic, git-diff-driven. Best for "small surgical edit" workflows and for teams that want full provider portability.
Cline (formerly Claude Dev, now under Q) — VS Code-extension-and-CLI hybrid. Cline became its own brand after the original Claude Dev rename; the tool itself is mature.

Behind the top five, the Pinggy 2026 round-up names opencode (75+ provider support), Qwen Code (Alibaba's CLI), and Cursor Agent (Cursor's CLI mode separate from the IDE) as the next tier. Goose (Block) and Warp 2.0 (its own category — see /topic/warp-2-agentic-terminal) round out the tools developers actually run in production.

That's nine tools in active production use, with another 20+ smaller projects mapped in the DEV comparison. The wave is wider than any one tool's marketing suggests.

Part 2: why CLIs and not IDEs

Three structural reasons the wave moved to the terminal, none of them aesthetic:

CLIs compose. A pipe between two CLIs is a deterministic, well-typed handoff. An agent that runs git status | head -20 | grep modified knows exactly what it gets back. The equivalent in an IDE — "ask the IDE to find modified files and pipe them to a subagent" — requires a custom RPC bridge for each pair of tools. The terminal already solved this in 1973. CLIs run anywhere. Remote SSH, CI runners, Docker containers, ephemeral sandboxes, cloud development environments. An AI coding CLI that's a binary plus a config file installs in 30 seconds in any of these. An IDE has to be packaged, shipped, updated, and authenticated separately in each. The economics favor the CLI. CLIs are auditable. Every byte the agent reads and every byte it writes is text in a scrollback buffer. You can grep for it, screenshot it, replay it. The IDE's rich event stream (file watchers, debounced edits, autosave, refactor-rename refactors) is a lot harder to audit when something goes wrong. For agent work — where "what did the model actually do" is the central question — auditability wins.

The standardization on CSI 2026 synchronized output (see /topic/csi-2026-synchronized-output) is the technical capstone: the terminal can now render rich, flicker-free, atomic UI updates, so the surface that was "good enough for shell scripts" is now also good enough for full agent dashboards.

Part 3: the second-order wave — platform CLIs got upgraded

The most interesting part of the trend isn't the AI tools themselves. It's that the platform CLIs — the ones from Stripe, Supabase, Vercel, PostHog, ElevenLabs, Ramp, Resend, Valyu — all shipped major updates in 2026 to be agent-friendly.

What "agent-friendly" means in practice, drawing from the DEV comparison's documentation of the second-order wave:

Structured JSON output by default. --json flags moved from afterthought to first-class. The agent doesn't have to parse English error messages anymore.
Predictable error codes. Exit codes mean specific failure classes (4 for auth, 5 for not found, 6 for rate limit). Agents can branch on them.
Idempotent operations. Running the same command twice produces the same result. Agents can retry without consequences.
Scoped tokens with narrow blast radius. Stripe's restricted keys, Supabase's role-scoped service keys, Vercel's project-scoped tokens. If the agent goes off the rails, the damage is bounded.
Help text written for agents. Stripe's CLI help now includes machine-readable examples and JSON schemas. The tooling assumes a non-human is reading.

This is the "tools updated to serve their callers" half of the wave, and it's why "CLI-everything" is a fair description rather than hyperbole. The agent ecosystem and the platform ecosystem are co-evolving toward a shared CLI standard.

How to choose: the honest decision tree

Most posts on this topic offer 20 recommendations and a comparison matrix. Here's the version that's actually useful:

What's your model provider? If it's Anthropic, install Claude Code. If it's OpenAI, install Codex CLI. If it's Google, install Gemini CLI. The first-party CLIs are the lowest-friction starting point.
Do you want provider portability? If yes, install Aider or opencode. Both work across providers, both ship features like git-diff-driven editing and multi-file refactor that the first-party CLIs sometimes lag on.
Is your work shell-heavy? If you spend more than 4 hours a day in a terminal, evaluate Warp 2.0 as a full terminal replacement. The single-surface design pays off when your other tools all run in the shell.
Are you running multiple agents in parallel? Install Conductor (Mac) or oh-my-hermes (cross-platform). The dual-wield workflows assume both Claude Code and Codex CLI are installed.

Most teams end up with 2-3 CLIs installed. The "pick one" framing is unrealistic; the realistic framing is "pick the one you use most, install one or two more for specific situations."

Where this fails / what we don't know

The 30+ count is a community map, not an industry census. The DEV community comparison is the best-published list, but it includes tools with very different maturity levels. Some are toys; some are funded products.
Vendor benchmarks are vendor benchmarks. Every CLI in this space has published a SWE-bench number that puts it favorably. Independent reproductions are sparse. Treat all "X beats Y" claims with skepticism until you've run both on your own codebase.
The platform-CLI upgrade story is uneven. Stripe and Vercel are far ahead. Smaller platforms are still catching up. The "all platform CLIs are agent-ready" framing oversimplifies.
No public daily-active-user data for most CLIs. Star counts and HN mentions are the proxies we have. Real DAU numbers would tell a different story; we don't have them.

Sources

DEV Community, "Every AI Coding CLI in 2026: The Complete Map (30+ Tools Compared)" — dev.to. Primary source for the 30+ count and the second-order CLI upgrade wave.
Pinggy, "Top CLI Based AI Coding Agents in 2026" — pinggy.io/blog. Top-5 ranking: Claude Code, Codex CLI, Gemini CLI, OpenCode, Aider.
Engineered Intelligence Substack, "Dual-Wielding Codex and Claude Code" — engineeredintelligence.substack.com. "2026 power stack" framing.
Warp blog, "Reimagining coding: agentic development environment" — warp.dev/blog. Warp 2.0 ADE category.
shakacode/claude-code-commands-skills-agents, "Claude Code with Codex" doc — github.com/shakacode. Canonical multi-CLI workflow doc.
HERMESquant/oh-my-hermes — github.com/HERMESquant/oh-my-hermes. The "omh handoff" multi-CLI orchestrator.
VS Code blog, "Multi-Agent Development" — code.visualstudio.com/blogs. GitHub Copilot v1.109 running Claude and Codex side by side.

Related GitHub projects

@anthropics

claude-code

Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.

⭐ 122,880

@Aider-AI

aider

aider is AI pair programming in your terminal

⭐ 44,722

@cline

cline

Autonomous coding agent right in your IDE, capable of creating/editing files, executing commands, using the browser, and more with your permission every step of the way.

⭐ 61,684

@anomalyco

opencode

The open source coding agent.

⭐ 159,149

@aaif-goose

goose

an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM

⭐ 45,093

Frequently asked

What is the 'CLI-everything' wave?: It's the 2026 pattern where AI coding tooling moved out of GUI-first IDEs and into the terminal. The wave has two parts: first-party AI coding CLIs (Claude Code, Codex CLI, Gemini CLI, Aider, Cline, Warp, opencode, Goose, Qwen Code, Cursor Agent, and 20+ more), and second-order platform CLIs (Stripe, Supabase, Vercel, PostHog, ElevenLabs, Resend) that rebuilt their CLIs to be agent-aware so the AI tools above can use them as deterministic action surfaces.
Why are AI tools moving to the CLI instead of staying in IDEs?: Three reasons that matter. First, CLIs compose — pipes, scripts, and exit codes are deterministic ways for one agent to call another. Second, CLIs run anywhere — remote SSH, CI, Docker, ephemeral sandboxes — without packaging an Electron app. Third, CLI surfaces give agents a stable text-only interface where every token of input and output is auditable, which the IDE's rich event stream doesn't offer.
How many AI coding CLIs are there in 2026?: The most comprehensive published comparison maps 30+ AI coding CLIs in 2026. The five most-discussed are Claude Code (Anthropic), Codex CLI (OpenAI), Gemini CLI (Google), Aider (open-source), and Cline (formerly Claude Dev — now Q, the company). Warp 2.0 is in a category of its own as the first 'agentic terminal' rather than a CLI tool you run inside a terminal.
Why did Stripe and Vercel update their CLIs in 2026?: The platform CLIs (Stripe, Supabase, Vercel, PostHog, ElevenLabs, Ramp, Resend, Valyu) updated their CLIs in 2026 to make them more agent-readable: structured JSON output, predictable error codes, idempotent operations, scoped tokens. The agent-to-CLI handoff is now a first-class integration surface, not an afterthought. This is the 'second-order wave' — AI coding agents need deterministic tools, so the tools themselves got upgraded.
Which AI coding CLI should I install first?: If you're on Anthropic, install Claude Code. If you're on OpenAI, install Codex CLI. If you want vendor-agnostic, install Aider or opencode (75+ providers). If your work is mostly shell + scripts, evaluate Warp 2.0 as a full terminal replacement. The honest answer is that the differences between the top 4-5 CLIs are smaller than the differences between any of them and an IDE-first workflow. Pick one, ship for a week, then decide.
Does RuleSell work with all 30+ AI coding CLIs?: RuleSell's catalog is designed around the cross-tool Markdown-rules pattern that most of these CLIs read. A well-written ruleset typically works in Claude Code, Cursor, Cline, Aider, Codex CLI, Gemini CLI, Warp, Windsurf, Continue, and other tools that accept Markdown agent config — with minor format adaptation per tool. The catalog explicitly markets to multi-CLI workflows, and creators publish once for the whole stack.