The CLI Is Back, But Weirder

Agentic CLIs like Claude Code, Codex CLI, and Antigravity CLI are nothing like the command-line tools you learned in a tutorial. Here's what changed, why it matters, and how to start using them.

TL;DR: The command line never went away, but it just got genuinely strange. A new wave of agentic CLIs - Claude Code, OpenAI Codex CLI, Google's Antigravity CLI - don't return text for you to copy. They read your codebase, plan a sequence of steps, run commands, check results, and iterate. That loop changes what you can build and how fast you can build it.

What the classic CLI actually does

If you've ever typed git commit or npm install, you've used a CLI - a command-line interface. You give it a specific command, it does one specific thing, it stops. That's the whole contract. git doesn't wonder what you're trying to accomplish. npm doesn't read your code and suggest a better package. They follow instructions, one step at a time, exactly as written.

This model has been around since the 1960s and it still works. Git, curl, ffmpeg, docker - these are precise, composable, reliable tools. The CLI's reputation for being hard to use mostly comes from the fact that you have to speak its exact language. Forget a flag and nothing happens, or worse, the wrong thing happens.

That's the baseline. What's happening now is something categorically different.

What "agentic" actually means

An agentic system acts toward a goal with a degree of autonomy, rather than responding to one prompt at a time. The word sounds like buzzword soup until you see the difference in practice.

Here's a traditional CLI interaction:

$ grep -r "TODO" ./src
./src/auth.ts:42: // TODO: handle token refresh
./src/api.ts:87: // TODO: add rate limiting

It found the TODOs. That's it. What you do next is entirely up to you.

Here's an agentic CLI interaction:

You: fix the token refresh TODO in auth.ts and write a test for it

Claude: I'll look at the auth flow first...
[reads auth.ts, reads session.ts, checks existing test patterns]
I see you're using JWT with a 15-minute expiry. Here's my plan:
1. Add a refreshToken() function in auth.ts
2. Call it when the API returns 401
3. Write a test that simulates an expired token

Shall I proceed?

The agentic CLI read files on its own, formed a plan, and asked for confirmation before touching anything. That's the shift: from "do this one thing I described exactly" to "work toward this goal, figure out the steps, check in when needed."

Anthropic describes this precisely in the Claude Code docs: Claude Code "reads your codebase, edits files, and runs commands across your terminal, IDE, desktop app, and browser." It doesn't wait for you to specify every step. It explores, plans, implements, and verifies.

The three tools you'll actually encounter

There are a few agentic CLIs worth knowing right now. They all do roughly the same thing - multi-step terminal work driven by natural language - but each has a different feel and set of tradeoffs.

Claude Code (Anthropic)

Claude Code is Anthropic's agentic coding environment. You install it, open a terminal in your project directory, and start describing what you want. It reads your files, plans multi-file changes, runs tests, commits code, and opens pull requests - all from the terminal.

What makes it different from just chatting with an AI: it runs in your local environment with real tools. It can call git, run your test suite, execute shell scripts, and connect to external services through the Model Context Protocol (MCP). It's not answering questions about code - it's writing and running code.

Claude Code also supports "auto mode," where a background AI classifier reviews each tool call before it runs. The classifier checks for three categories of risk: scope escalation (is Claude doing something beyond what you asked?), untrusted infrastructure (is the action targeting systems you haven't sanctioned?), and prompt injection (does the action look like it was driven by hostile content Claude found in a file or web page?). Actions that clear all three checks proceed without interrupting you. Anything that doesn't gets blocked and surfaces for your review.

One quirk worth knowing: Claude Code manages a context window - a running log of your conversation and everything it has read. That window fills up. Longer sessions can degrade performance as the window gets crowded. The best practice from Anthropic's own docs: run /clear between unrelated tasks, and give Claude a way to verify its own work (a test suite, a build check) so it can close the feedback loop without you watching every step.

Codex CLI (OpenAI)

OpenAI's Codex CLI is open source (Apache 2.0) and runs locally in your terminal. By default it's in "suggest mode" - every file edit and every command gets proposed before execution. Nothing runs without explicit approval. It's the safest starting point if you're nervous about an AI with terminal access.

Codex also has scripting mode via the exec command, which lets you automate repetitive workflows without a human in the loop. It supports MCP for extending what it can connect to, and you can pass images directly into prompts - useful if you're building from a design mockup.

Antigravity CLI (Google)

Google originally released Gemini CLI as an open-source terminal agent in June 2025. In May 2026, Google launched Antigravity CLI - a rebuilt version written in Go, positioned as an "agent-first development platform." The new tool starts faster, uses less memory, and supports asynchronous multi-agent workflows, where background tasks can run in parallel while you keep working in the foreground.

Free-tier users of the old Gemini CLI needed to migrate by June 18, 2026. Enterprise customers on Gemini Code Assist Standard or Enterprise licenses kept their existing access unchanged. Antigravity CLI ties into the same Gemini models but with a unified architecture shared across Google's desktop and web tools.

Why this is different from "just asking ChatGPT"

This is the question worth sitting with if you want to actually use these tools rather than just nod at them.

A chat interface gives you text. You have to read that text, decide what to do with it, copy parts of it into files or your terminal, run commands yourself, notice if something went wrong, and go back for more help. Every step requires you.

An agentic CLI closes that loop. It reads files, it edits files, it runs commands, it reads the output of those commands, it decides what to do next. You're in the loop as an approver and redirector - not as the person who copies the answer out of a chat window.

One concrete way to feel this: give Claude Code a bug with a stack trace. Don't explain the bug. Just paste the error. It will read the relevant files, find the likely source, propose a fix, and ask you to confirm. You're approving and guiding, not transcribing.

Another: give it a GitHub issue number. It will read the issue, scan the relevant code, plan the implementation, write it, run the tests, and open a PR. The whole software development loop from issue to PR can close without you switching windows.

The part where you still have to pay attention

Agentic CLIs are genuinely powerful and genuinely capable of doing things you didn't intend. A tool that can run shell commands and edit files in your project can also make a mess of your project if you give it an ambiguous goal and walk away.

A few things that actually matter in practice:

The underlying plumbing: MCP

All three major agentic CLIs support the Model Context Protocol - an open standard that Anthropic introduced in 2024 and donated to the Agentic AI Foundation (under the Linux Foundation) in December 2025. As of early 2026, MCP has over 97 million monthly SDK downloads and support from every major AI vendor.

MCP is how agentic CLIs connect to the world outside your file system. Instead of each tool hard-coding integrations with GitHub, Notion, Stripe, Figma, and every other service you use, MCP lets any service expose a standard interface that any MCP-compatible tool can consume. You install an MCP server for a service once, and your agentic CLI can use it.

The practical implication: Claude Code with a Stripe MCP server can look at a failing payment and the relevant code in the same session. It's not switching tabs. It's not asking you to copy an error from one window to another. It has context from both places at once.

This is what makes agentic CLIs feel genuinely different from a smarter autocomplete - they're connected to the actual systems your software runs on.

Try this next: Once you're comfortable with what agentic CLIs can do on their own, the natural next step is connecting them to the services your project actually uses. Read MCP: Connecting Your Tools to Any AI Agent to see how to wire up GitHub, databases, and other services so your CLI has full context - not just your local files.

LearntoolkitThe CLI Is Back, But Weirder
Guidetoolkitintro7 min read

The CLI Is Back, But Weirder

Agentic CLIs like Claude Code, Codex CLI, and Antigravity CLI are nothing like the command-line tools you learned in a tutorial. Here's what changed, why it matters, and how to start using them.

CREATORS OF TODAY.
@creatorsoftoday
321

TL;DR: The command line never went away, but it just got genuinely strange. A new wave of agentic CLIs - Claude Code, OpenAI Codex CLI, Google's Antigravity CLI - don't return text for you to copy. They read your codebase, plan a sequence of steps, run commands, check results, and iterate. That loop changes what you can build and how fast you can build it.

What the classic CLI actually does

If you've ever typed git commit or npm install, you've used a CLI - a command-line interface. You give it a specific command, it does one specific thing, it stops. That's the whole contract. git doesn't wonder what you're trying to accomplish. npm doesn't read your code and suggest a better package. They follow instructions, one step at a time, exactly as written.

This model has been around since the 1960s and it still works. Git, curl, ffmpeg, docker - these are precise, composable, reliable tools. The CLI's reputation for being hard to use mostly comes from the fact that you have to speak its exact language. Forget a flag and nothing happens, or worse, the wrong thing happens.

That's the baseline. What's happening now is something categorically different.

What "agentic" actually means

An agentic system acts toward a goal with a degree of autonomy, rather than responding to one prompt at a time. The word sounds like buzzword soup until you see the difference in practice.

Here's a traditional CLI interaction:

$ grep -r "TODO" ./src
./src/auth.ts:42: // TODO: handle token refresh
./src/api.ts:87: // TODO: add rate limiting

It found the TODOs. That's it. What you do next is entirely up to you.

Here's an agentic CLI interaction:

You: fix the token refresh TODO in auth.ts and write a test for it

Claude: I'll look at the auth flow first...
[reads auth.ts, reads session.ts, checks existing test patterns]
I see you're using JWT with a 15-minute expiry. Here's my plan:
1. Add a refreshToken() function in auth.ts
2. Call it when the API returns 401
3. Write a test that simulates an expired token

Shall I proceed?

The agentic CLI read files on its own, formed a plan, and asked for confirmation before touching anything. That's the shift: from "do this one thing I described exactly" to "work toward this goal, figure out the steps, check in when needed."

Anthropic describes this precisely in the Claude Code docs: Claude Code "reads your codebase, edits files, and runs commands across your terminal, IDE, desktop app, and browser." It doesn't wait for you to specify every step. It explores, plans, implements, and verifies.

The three tools you'll actually encounter

There are a few agentic CLIs worth knowing right now. They all do roughly the same thing - multi-step terminal work driven by natural language - but each has a different feel and set of tradeoffs.

Claude Code (Anthropic)

Claude Code is Anthropic's agentic coding environment. You install it, open a terminal in your project directory, and start describing what you want. It reads your files, plans multi-file changes, runs tests, commits code, and opens pull requests - all from the terminal.

What makes it different from just chatting with an AI: it runs in your local environment with real tools. It can call git, run your test suite, execute shell scripts, and connect to external services through the Model Context Protocol (MCP). It's not answering questions about code - it's writing and running code.

Claude Code also supports "auto mode," where a background AI classifier reviews each tool call before it runs. The classifier checks for three categories of risk: scope escalation (is Claude doing something beyond what you asked?), untrusted infrastructure (is the action targeting systems you haven't sanctioned?), and prompt injection (does the action look like it was driven by hostile content Claude found in a file or web page?). Actions that clear all three checks proceed without interrupting you. Anything that doesn't gets blocked and surfaces for your review.

One quirk worth knowing: Claude Code manages a context window - a running log of your conversation and everything it has read. That window fills up. Longer sessions can degrade performance as the window gets crowded. The best practice from Anthropic's own docs: run /clear between unrelated tasks, and give Claude a way to verify its own work (a test suite, a build check) so it can close the feedback loop without you watching every step.

Codex CLI (OpenAI)

OpenAI's Codex CLI is open source (Apache 2.0) and runs locally in your terminal. By default it's in "suggest mode" - every file edit and every command gets proposed before execution. Nothing runs without explicit approval. It's the safest starting point if you're nervous about an AI with terminal access.

Codex also has scripting mode via the exec command, which lets you automate repetitive workflows without a human in the loop. It supports MCP for extending what it can connect to, and you can pass images directly into prompts - useful if you're building from a design mockup.

Antigravity CLI (Google)

Google originally released Gemini CLI as an open-source terminal agent in June 2025. In May 2026, Google launched Antigravity CLI - a rebuilt version written in Go, positioned as an "agent-first development platform." The new tool starts faster, uses less memory, and supports asynchronous multi-agent workflows, where background tasks can run in parallel while you keep working in the foreground.

Free-tier users of the old Gemini CLI needed to migrate by June 18, 2026. Enterprise customers on Gemini Code Assist Standard or Enterprise licenses kept their existing access unchanged. Antigravity CLI ties into the same Gemini models but with a unified architecture shared across Google's desktop and web tools.

Why this is different from "just asking ChatGPT"

This is the question worth sitting with if you want to actually use these tools rather than just nod at them.

A chat interface gives you text. You have to read that text, decide what to do with it, copy parts of it into files or your terminal, run commands yourself, notice if something went wrong, and go back for more help. Every step requires you.

An agentic CLI closes that loop. It reads files, it edits files, it runs commands, it reads the output of those commands, it decides what to do next. You're in the loop as an approver and redirector - not as the person who copies the answer out of a chat window.

One concrete way to feel this: give Claude Code a bug with a stack trace. Don't explain the bug. Just paste the error. It will read the relevant files, find the likely source, propose a fix, and ask you to confirm. You're approving and guiding, not transcribing.

Another: give it a GitHub issue number. It will read the issue, scan the relevant code, plan the implementation, write it, run the tests, and open a PR. The whole software development loop from issue to PR can close without you switching windows.

The part where you still have to pay attention

Agentic CLIs are genuinely powerful and genuinely capable of doing things you didn't intend. A tool that can run shell commands and edit files in your project can also make a mess of your project if you give it an ambiguous goal and walk away.

A few things that actually matter in practice:

  • Approval modes exist for a reason. Start with the most cautious mode - Codex's suggest mode, Claude Code's default permission prompts. Watch what it proposes before you trust it to run autonomously.
  • Give it a way to verify its own work. A test suite it can run is worth more than your approval of every individual file edit. If it can check itself, sessions you're not watching can still converge on correct output.
  • Scope your tasks. "Make the app better" is a bad prompt for any tool. "Fix the 401 handling in auth.ts and add a test" is a good one. Narrow scope means fewer surprises.
  • Check what it actually did. git diff before you commit. These tools work fast and change multiple files simultaneously - reviewing the diff is faster than reviewing each file individually.

The underlying plumbing: MCP

All three major agentic CLIs support the Model Context Protocol - an open standard that Anthropic introduced in 2024 and donated to the Agentic AI Foundation (under the Linux Foundation) in December 2025. As of early 2026, MCP has over 97 million monthly SDK downloads and support from every major AI vendor.

MCP is how agentic CLIs connect to the world outside your file system. Instead of each tool hard-coding integrations with GitHub, Notion, Stripe, Figma, and every other service you use, MCP lets any service expose a standard interface that any MCP-compatible tool can consume. You install an MCP server for a service once, and your agentic CLI can use it.

The practical implication: Claude Code with a Stripe MCP server can look at a failing payment and the relevant code in the same session. It's not switching tabs. It's not asking you to copy an error from one window to another. It has context from both places at once.

This is what makes agentic CLIs feel genuinely different from a smarter autocomplete - they're connected to the actual systems your software runs on.

  • Traditional CLIs follow exact commands, one step at a time. Agentic CLIs pursue goals across multiple steps, reading and writing your actual files along the way.
  • The big three right now: Claude Code (Anthropic), Codex CLI (OpenAI, open source Apache 2.0), Antigravity CLI (Google, formerly Gemini CLI, rewritten in Go).
  • The core loop is: read context, plan, act, verify, iterate. You stay in the loop as an approver, not as the person doing the mechanical steps.
  • MCP is the plumbing that connects these tools to your services - databases, GitHub, Stripe, Figma - through a single open standard, now governed by the Agentic AI Foundation under the Linux Foundation.
  • Start cautious: use approval modes, give the tool a way to verify its own work, review the diff before committing.
  • Agentic CLIs don't replace your judgment. They eliminate the time between having a plan and having code that executes it.

Try this next: Once you're comfortable with what agentic CLIs can do on their own, the natural next step is connecting them to the services your project actually uses. Read MCP: Connecting Your Tools to Any AI Agent to see how to wire up GitHub, databases, and other services so your CLI has full context - not just your local files.

References & sources

Reviews

Only verified humans can leave reviews. It keeps every rating real.

Verify to review

No reviews yet. Be the first to share your take.