Agentic capabilities are the next major evolution in using LLMs to aide in software development work. Now you can simply open a folder, tell it what to build, and the model will go off and do it. The model will generate a to-do list, run commands, search the web, compile code, and iterate through issues completely independently until its task is done. At least thats the ideal use case. Here I just want to compare all of the tools I’m using that enable these agentic capabilities.
opencode
This is the only truly Platform Agnostic and Open Source tool that I’ve used so far. You can use your own api key from anything that is listed on https://models.dev (which is exhaustive in my opinion), and even run locally with tools like Ollama.
Being open source and fully independent definitely comes with its rough edges. I can’t use the tool with ChatGPT-5 because it hits rate limiting, my coworkers had their install totally kill itself, and occasionally it just gets stuck. However, the dev community is lively and if you report a bug someone at least looks at it.
Codex
This is the offering from OpenAI. Its technically Open Source but we all know how much they actually value keeping things open. This is by far the most barebones tool, but it has a few features I really like. I like that theres a few commands like git status that I can run from inside of the tool so that I don’t have to open another ghostty tab.
Claude Code
Super powerful, I just don’t like the direction Anthropic is taking with vendor lock in. You must have a anthropic account linked, you can’t just use an API key like even Codex supports. This was a major pain for my work, and I could see it being an issue for anyone using AWS Bedrock models. They also really want you to use Claude.md instead of AGENTS.md in your projects, and things like the memories feature seem pretty locked in.
All this to say I think its the highest quality tool, I definitely get the best results out of it.
Cursor
This is the only app here that isn’t a TUI. This is just a fork of VS Code that has some opinionated changes that I really like. Cursor is surprisingly the worst at vendor lock in. You must have a subscription to use the editor, you can bring your own keys but they try really hard to deter you from doing so.
There are some major advantages to being a full code editor for working alongside a LLM like giving it context on what files are open, quickly sharing terminal errors, and having a better UI for seeing changes.
I do find the app very buggy though. Its pretty often when I give it contexts through the UI instead of seeing my terminal it just sees @terminal so the find and replace isn’t working as intended. I also think their prompts are the worst out of all the tools I use. Models used from cursor are incredibly chatty and often don’t complete tasks.