Aman's Musings

AI Tools I am using in 2025

This year I used a bunch of AI tools for both work and personal things. These are my thoughts on what I like and dislike about each of them. Since I work at AWS, there is an obvious bias toward Amazon products and Anthropic models because those are the tools I use most internally.

LibreChat

I deployed LibreChat on my Pi that sits on my private Tailscale VPN. I use it as my ChatGPT replacement across all devices. Switching between different models feels from a single app feels very natural and it works well for exploration and thinking tasks. For coding and thinking related queries, I usually rely on Claude, and for everything else I fall back to GPT or Gemini.

I also plan to integrate LibreChat with the AWS Bedrock Prompt Router. My idea is to use Amazon Nova for prompt routing to a particular model because it is both cheap and good enough for this type of workload. The goal is to automatically route conversations to the right model and keep my API bills frugal without sacrificing response quality.

Q CLI

This is the tool I spent the most time with. It is built in Rust, which already earns bonus points for me because I love coding in rust. I have followed this project since the time it was still called Fig before Amazon acquired it. It originally started as a terminal autocomplete tool similar to VS Code IntelliSense and eventually evolved into the agentic AI CLI that it is today.

I really like how simple and clean the tool feels while still being powerful, especially with the latest Claude models. The new agent feature is genuinely useful because it lets me shift between different SDE modes. I can go from writing Java to working with CDK to debugging pod issues on K8s clusters without friction.

The free tier is generous, so I even use it for personal things. I often talk to my Pi with it, manage dotfiles and handle random tasks on my Mac Mini. I built an agent to SSH into my Pi using natural language and it turned out to be extremely handy. Gemini CLI refused to perform SSH because it considered it too risky, so this was a nice workaround.

The MCP server support also works well. At this point everything seems to have an MCP server, but using the AWS CLI MCP server with Q CLI genuinely improves my workflow. I rarely open the AWS Console anymore and can stay inside iTerm2 for almost all tasks. Since I am part of the internal dev channel, I get to see the ongoing development and beta releases and I am excited about the direction the tool is taking.

Kiro

This year was full of VS Code and Chromium forks like Cursor, Windsurf, Comet and Atlas. Kiro joined the list toward the end of the year. Since I wrote a lot of non-Java code and had moved out of IntelliJ, VS Code became my main editor, so trying Kiro was natural.

I used it here and there for quick POCs and scripting. The project is still in a very early state and things break occasionally, but I have had decent success when starting with fresh repositories. The spec driven workflow helps bring vague ideas into a structured list of actual tasks. The task queue is useful too, although I would really like a DAG style system so tasks can run in parallel. The hooks feature is neat because it automates things like linting and git checkpoints.

Right now my main problem is that it struggles with large established codebases. Context gathering in vibe mode is not strong enough, which pushes me back to the Amazon Q extension for VS Code or to Cline whenever I need more pointed reasoning. Kiro also has a habit of over engineering things. It creates unnecessary files and random markdown docs for very small tasks. The tool aims to provide an AI experience for enterprise code, but sometimes feels too eager to behave like an enterprise product even when that is not required.

Still, the shipping velocity is fast and I am hopeful it will mature quickly.

Cline

I mostly use Cline in VS Code. It is reliable and does a good job when you want to provide very specific context such as a few files and a clear instruction. The fact that it is open source and allows custom API keys is perfect when I feel like experimenting with models other than Claude like GPT OSS or LLaMA 70B.

It also gives better visibility into what the model is doing. I can see the model thinking, the XML style prompts, token costs and context window usage. This makes it easy to steer and avoids hallucinations caused by bloated context. The browser integration is quite helpful whenever I work on frontend tasks.

PartyRock

PartyRock is simple and it always works. I mostly use it for lightweight CustomGPT style workflows such as proofreading or adjusting tone. One of my favorite mini workflows is a follow up message generator. I give it the original text, the time since I sent it and the urgency level. It generates a well balanced follow up that almost always gets me a faster response.

Cursor

Cursor is the state of the art editor right now, but I have barely used it. It evolved a lot with the Cursor 2 release and I need to spend more time with it before forming an opinion. I am also excited to see their new Composer coding models and compare its quality with Claude Sonnet series of model which are my favorite models for coding tasks.

Claude Code

I recently got the Claude subscription and it absolutely lives up to the hype. Since I prefer spending time inside iTerm2 rather than VS Code or its forks, Claude Code has quickly become my favorite AI coding tool.

The TUI feels extremely comfortable. Many features overlap with Q CLI, but the experience feels smoother overall. I even like the random words it shows while loading responses. It is not a functional feature but it looks nice. I genuinely feel Anthropic is pushing the envelope with this product. Right now it feels like the best AI coding experience available.

Gemini CLI

I tried it mainly because it is cheap. It feels a bit like a clone of Claude Code. Since I prefer Claude models anyway, I did not end up using it much.

Perplexity

This is my default search engine. The finance section is excellent and I honestly think it could surpass Yahoo Finance in user experience at some point. I use its schedule and alert feature regularly to track stock prices and related news.

Wispr Flow

I use Wispr Flow on my personal machine and it is great when I need to write without typing. I also tried writing python code through voice and it worked better than I expected.

At work I tried deploying OpenAI Whisper locally using the menu bar tool I fork-built https://github.com/apsknight/whisper-dictation. I eventually stopped because it felt awkward to talk to my laptop in an open office while pleading Q CLI to fix my code.

Zoom AI Meeting Summary

The accuracy of the summaries is surprisingly high even when based only on voice. I do suspect it also uses screen share snapshots but the results are still impressive. It is extremely helpful on days when I zone out and need a clean recap of what was discussed along with action items.

Thoughts on MCP

I used MCP everywhere earlier, but over time I noticed it polluted the context window. The tool calls were too verbose. The recent Anthropic blog post on this topic pretty much echoed everything I was already thinking - https://www.anthropic.com/engineering/code-execution-with-mcp

Letting the model invoke command line tools directly or write and execute code tends to be much more efficient. These days I keep only a few essential MCP servers and let the model handle the rest locally by writing custom code so the context window stays clean.

I also created a small tool to quickly toggle the installed MCP servers so that I can easily enable/disable the required MCP servers whenever I want - https://github.com/apsknight/mcp-manager

Tools I Want to Try

These are some tools I am planning to try soon.