Last year, every AI coding tool was racing to add more integrations, more plugins, more tool definitions. Now a counter-argument is gaining traction: maybe a terminal and a filesystem is all an AI agent actually needs.
The core claim, laid out in a recent analysis by developer Nishant Soni and backed by a ServiceNow research paper titled "Terminal Agents Suffice for Enterprise Automation," is straightforward. Every major LLM was trained on enormous amounts of shell content: man pages, Stack Overflow answers, GitHub repos, sysadmin guides. The terminal is probably the single most well-represented human-computer interface in LLM training data. When you give a model a shell, you're meeting it where it already has deep competence.
The Intelligence Tax of Tool Sprawl
Custom tool integrations - including the increasingly popular MCP (Model Context Protocol) standard - require the model to learn unfamiliar abstractions from scratch via context window descriptions. Every tool definition, schema, and usage instruction stuffed into the context window is cognitive budget the model can't spend on your actual problem.
The argument isn't theoretical. The ServiceNow AI team found that terminal-only agents matched or outperformed MCP-augmented architectures across real-world enterprise tasks. The reasoning: every API can be called with curl. Every file can be read with cat. Every transformation can be handled with standard Unix utilities or a quick script. You don't need a custom Jira tool or a Salesforce wrapper. You need the API docs and a shell.
This maps to what Claude Code, Aider, and Cline users already experience daily. The most capable AI coding sessions tend to be the ones where the agent writes and executes code directly rather than routing through layers of tool abstractions.
Where This Breaks Down
The terminal-first argument has real limits. Non-technical users can't debug a failed curl command. GUI-heavy workflows like design tools or spreadsheet manipulation don't reduce cleanly to shell commands. And MCP's value isn't zero - it provides structured, discoverable interfaces that reduce the chance of the model hallucinating an API endpoint or misremembering a flag.
The practical takeaway for anyone choosing an AI coding setup: if you're a developer comfortable in the terminal, resist the urge to install every MCP connector and plugin available. Start with the shell. Add integrations only when you hit a genuine wall. The model is almost certainly better at bash than it is at navigating your custom tool definitions.
Soni, it should be noted, is building a product called nonbios based on this exact thesis - a terminal agent with zero tools and no MCP. That's a clear conflict of interest, but the ServiceNow research stands on its own.