AI coding tools pivot toward development in the terminal

For years, AI-enhanced code editors like Cursor, Windsurf, and GitHub Copilot have set the standard for software development. However, as agentic AI grows more capable and a more intuitive "vibe-coding" approach gains traction, a quiet evolution is reshaping how these systems operate. Rather than solely manipulating code, they are increasingly engaging directly with the system's command-line shell. This marks a fundamental shift in AI-assisted development—a subtle but potentially transformative move for the industry's future direction.

The terminal, often reminiscent of the stark black-and-white screens from 90s hacker films, is a classic interface for executing programs and managing data. While less visually polished than modern code editors, it remains an immensely powerful tool in skilled hands. Although code-centric agents can write and debug programs, terminal-based utilities are frequently essential for turning written code into a functional, deployable application.

The move toward the terminal is most evident from leading AI labs. Since February, Anthropic, DeepMind, and OpenAI have each launched command-line coding tools (Claude Code, Gemini CLI, and CLI Codex, respectively), which have quickly become some of their most popular offerings. This transition is easy to overlook, as these tools often share branding with their predecessors. Yet, fundamentally, how AI agents interact with computers—both online and off—has changed. Many believe this is just the beginning.

"Our core hypothesis is that in the future, roughly 95% of LLM-computer interaction will occur through a terminal-like interface," says Alex Shaw, co-creator of the prominent terminal-focused benchmark, TerminalBench.

Terminal-based tools are gaining prominence just as some established code-focused platforms face instability. The AI code editor Windsurf, for instance, has been fragmented by competing acquisitions, with senior leaders recruited by Google and the remaining company bought by Cognition—casting doubt on the consumer product's long-term viability.

Simultaneously, new research suggests developers might overestimate the productivity benefits of traditional tools. A METR study evaluating Cursor Pro, Windsurf's primary rival, found that while developers predicted a 20-30% speed increase, the observed process was nearly 20% slower. In essence, the code assistant was, in some cases, costing programmers time.

This has created an opportunity for companies like Warp, which currently leads the TerminalBench rankings. Warp describes itself as an "agentic development environment," bridging the gap between full IDEs and command-line tools like Claude Code. Founder Zach Lloyd remains optimistic about the terminal's potential, viewing it as a way to solve problems beyond the scope of a traditional code editor like Cursor.

LIVE NOW! TechCrunch All Stage

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and beyond for a day packed with strategies, workshops, and meaningful connections.

Save $450 on your TechCrunch All Stage pass

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and beyond for a day packed with strategies, workshops, and meaningful connections.

"The terminal sits at a very foundational layer in the developer stack, making it the most versatile place to run AI agents," Lloyd explains.

To grasp how this new approach differs, consider the benchmarks used to evaluate them. The previous generation of code-based tools was measured on solving GitHub issues, the foundation of the SWE-Bench test. Each SWE-Bench problem is an open GitHub issue—a piece of malfunctioning code. Models iteratively refine the code until it works. While integrated products like Cursor have developed more sophisticated methods, the core GitHub/SWE-Bench model remains: start with broken code and fix it.

Terminal-based tools adopt a broader perspective, considering the entire environment a program runs in, not just the code. This encompasses coding as well as DevOps-oriented tasks like configuring a Git server or diagnosing why a script fails. One TerminalBench challenge provides a decompression program and a target text file, tasking the agent with reverse-engineering a matching compression algorithm. Another asks the agent to build the Linux kernel from source, without mentioning it must first obtain the source code. Solving these requires the persistent, creative problem-solving intrinsic to programming.

"What makes TerminalBench difficult isn't just the questions we pose to the agents," says Shaw, "it's the complex environments we place them in."

Critically, this new method involves tackling problems through sequential, reasoned steps—the same capability that makes agentic AI so powerful. Yet, even cutting-edge agentic models struggle in some of these environments. Warp achieved its top TerminalBench score by solving just over half of the problems, highlighting both the benchmark's difficulty and the work still needed to realize the terminal's full potential.

Nevertheless, Lloyd contends we've reached a point where terminal-based tools can reliably manage a significant portion of a developer's non-coding workload—an offer that's difficult to refuse.

"Consider the daily work of setting up a new project, figuring out dependencies, and making it runnable—Warp can largely handle that autonomously," says Lloyd. "And if it can't, it will tell you exactly why."