My favourite approach is to use the terminal even to allow agents editing the code.

The tools I tried are:

• Claude Code CLI - From Anthropic, will use Claude API and requires some credit to use it
• Crush - Open source, can use most API providers and even your local AI from Ollama (in this case, results will be poor. Local AI can achieve decent results but not yet as agent)

The agent can

1. run command in your terminal to inspect files, try running code, etc.
2. edit multiple files to achieve a task

In the screenshot you see Claude Code, with my prompt highlighted, and then it’s suggesting to run a command and asking permission. Running the command, it will be able to see the output (error) and make a plan to identify the issue.

---