Claude Code Improvement Suggestions

Bug Description
Hilo señors and señoras from Anthropic. Below are some suggestions for things I noticed throughout my last few days using Claude Code and some suggestions I have for the tool. Some are cosmetic, some are more functional:

Cosmetic Suggestions

A way to set the color of the output from commands. Currently all output are just plain red, which sometimes feels like I ran into an error until I read it and realise that was not the case. Maybe allowing the base terminal color system when running commands like uv pip install would help.

When the Code agent is thinking/generating it’d be cool to see what the word/verb means as it’s thinking. Sometimes the Code agent can take up 300 seconds to complete a task/generate code and it’d be fun to have a little fact about the word while we wait. Maybe using Haiku 3 or even 3.5 to generate this would be nice and I personally wouldn’t mind paying a little extra for that. It can be a configurable setting as well so it’s off by default but can be enabled through the /config command

Maybe adding streaming support? I’m not sure how this would look considering Claude uses tools but it’d be nice to see some of the time.

Token usage/cost at the end of each message. Similar to aider, it allows users to see how much they’ve spent without needing to run /cost. Instead it’ll show a brief and simple token usage counter and cost per message + total cost so far for the session. Then in the /cost command maybe we get a more detailed view?

Functional Suggestions

1. Ability to expand Claude Code’s command capabilities with mcp. Essentially people can build mcp servers that allow you to use custom commands. Like maybe if I wanted a specific git workflow I could make a server that exposes a /commit command which in theory can add and commit my changes with a message without needing to enter bash mode. This can make Claude Code very extensible and customisable for everyone and allow people to potentially intgrate existing workflows into Claude code

1. Adding the ability to auto-compact. When the codebase is larger and/or the session just gets long, it’s be helpful to allow Claude to decide the best time to compact its own context. For example maybe it’s done with a task and before it moves on to the next task, it knows that the two tasks are not really similar in nature and need different pieces of context, well Claude can call an auto-compact tool to do it automatically. And Claude can even clear context with your approval. This would make it much easier to manage context and could allow Claude to decide “hey I need more space for the upcoming task and the previous things I did were not really related so let me clean this up a bit”

1. Add custom tools to Claude and specify which tools to use for specific code bases. I guess this can be done via MCP servers but I thought it can be somewhat separate. Maybe I have this custom way to to test whether my code is working via a script I have for all my Python projects. It’d be nice to add a schema and tell Claude in my Python project via the CLAUDE.md file to go “use this tool to test my code” and boom. Done.

1. I think it would be super cool to have Claude intelligently search for content rather than using the grep tool. Because I feel like the grep tool often adds to the large amount of context which can drive up costs.5. Add some other sub-agents. In aider we have an architect which is kind of like a planner and we’re able to discuss the plan without Claude or whatever LLM you’re using to actually edit your files. This would be helpful as sometimes I’m just asking Claude what are some options for optimising my codebase or script and it goes and tries to edit the files even though I more so wanted to hear some ideas and some pseudo-code perhaps.

And voila. These are my suggestions for how to make Claude Code better. Thank you for coming to my TED talk

Environment Info

• Platform: macos
• Terminal: WarpTerminal
• Version: 0.2.19
• Feedback ID: 4a2a0c06-ac56-4389-bcdc-7e70d9f1d0e5