Goodbye Claude, Hello Pi

I have been a longtime Claude Code user since it became available in the Claude Pro subscription plan over a year ago. I remember the first time I used it, feeling excited as it generated messy Python code from a single prompt for my various weekend projects. Despite this enthusiasm, I have now decided to retire Claude Code as my primary coding tool and migrate my development setup to use Pi instead.

On the day of my decision, Anthropic announced that they would finally be supporting AGENTS.md files in Claude Code, hoping to retain my loyalty. However, for me, it was too little, too late. I would like to take this opportunity to share my grievances with Claude Code and the reasons I am looking forward to avoiding them with Pi.

One of the main issues with Claude Code is the AGENTS.md feature. While it is great for guerrilla marketing, having numerous CLAUDE.md files scattered in various GitHub hobby repos can be cumbersome for users who wish to review the code before committing it. I prefer to manually review most of the code generated by Claude Code before integrating it into my projects.

Another problem I encountered was the limitation of having only one session connected to the IDE at a time when using Claude Code inside VSCode. This can be frustrating when working on multiple tasks simultaneously, as the second tab cannot receive text selection events from the IDE. This issue has been reported since December 2025 but was closed as a duplicate of an unrelated bug.

Furthermore, I have experienced numerous issues with the Claude Code VSCode GUI extension. Many built-in slash commands do not work, it lacks a custom status line, and subagent output messages are simply layered into the main chat thread, making it difficult to read when multiple sessions are running.

Despite these frustrations, I appreciate the rapid pace at which Anthropic has been adding features to Claude Code. Dynamic workflows, agent teams, and ultrareviews are impressive tools for enterprise customers, but they do not align with my personal needs. Additionally, the frequent changes to system prompts and tool descriptions have caused significant disruptions to my workflow, often contradicting user prompts or repository conventions.

While I acknowledge Anthropic's innovative efforts in the LLM coding harness space, I believe that a more focused approach would yield better results. The constant tinkering with features, such as dynamic workflows and agent teams, may seem impressive but ultimately do not cater to the needs of all users. I feel that there should be a dedicated engineer whose sole responsibility is to say "no" to certain requests, ensuring a more stable and user-friendly product.

Overall, Claude Code feels like a product without a clear vision or direction. It appears that a group of enthusiastic engineers and product managers are rapidly adding features based on their own workflows, without considering the potential drawbacks for other users. I believe that a strong product vision, coupled with careful decision-making, is crucial for success in this rapidly evolving field.

As I transition to using Pi as my primary coding harness, I am excited to embrace its stability, minimalist approach, and extreme customizability. The initial setup required some adjustments, but since then, I have found it to be a smooth experience. I have already developed a few extensions to support my specific use cases, which I plan to share once they are refined.

For those who share my frustrations with Claude Code, I highly recommend giving an open-source harness like Pi a try. It may seem daunting to implement custom solutions, but the benefits are worth the effort.