Developer switches from Claude Code to Pi for better stability

The move to open source

A longtime user of Claude Code recently decided to migrate their entire development setup to Pi, an open-source coding harness. While the user initially enjoyed the speed of generating Python code from single prompts, a series of small frustrations eventually led to the switch. This transition comes despite Anthropic's recent efforts to add features like AGENTS.md support to retain users.

The author describes Claude Code as a product that appears to lack a singular, clear vision. Instead, it seems to be a collection of features tacked on by engineers to suit specific workflows. This rapid development can result in a tool that feels cluttered and less stable for the average developer who needs a reliable daily driver.

In contrast, the user found Pi to be a much smoother experience after a brief initial setup period. They highlighted the stability and minimalist approach of the open-source tool as primary reasons for the switch. By moving to Pi, the developer can prioritize a tool that fits their specific needs rather than one shaped by corporate requirements.

Technical friction in Claude

One major pain point for the user involves the limitations of Claude Code's integration with VSCode. While the tool allows users to highlight text to include it in the AI's context, it only supports one active session at a time. This means a developer cannot work on two different tasks in separate tabs simultaneously while maintaining that context.

The author notes that while an official GUI extension exists to solve this, it introduces new problems. Half of the built-in slash commands do not function in the GUI, and the output messages for subagents are layered into the main chat. This makes it difficult to read the conversation when multiple tasks are running at once, creating a messy user interface.

Furthermore, the user expressed frustration with the constant updates to the system prompts. Because over 3/4 of the 244 releases this year modified these instructions, the model's performance can change unpredictably. These changes can sometimes contradict the user's own prompts or the established conventions of the specific code repository they are working on.

The cost of overhead

The developer pointed out significant inefficiencies in how Claude Code handles tool descriptions. Specifically, the default Bash tool description is almost 3,000 tokens long, much of which covers basic git and GitHub commands. The author argues that modern models already know how to perform these actions without such detailed, repetitive instructions.

From a technical standpoint, this overhead can have real financial implications for large-scale operations. The author suggests that if a company were processing 100 million requests per day, trimming 1,000 tokens could save millions of dollars in inference costs annually. Reducing this bloat would be a simple way to improve both performance and cost-efficiency for enterprise customers.

Despite these issues, the author acknowledges that Anthropic has created a very cool piece of software. They recognize that being an early innovator in the LLM coding space is difficult and requires significant trial and error. However, they believe that later tools like Pi have the benefit of hindsight to determine which features are essential and which can be removed.

Customization with Pi

One of the biggest advantages of switching to an open-source harness like Pi is the level of customizability it offers. The author was able to write a custom extension in about 20 minutes to handle functionality across multiple open tabs. This specific feature was a major limitation in the Claude Code environment that the user wanted to overcome.

The user also noted that Pi allows for more precise control over how the model interacts with the system. For example, they were able to instruct the model to use specific commands that bypass certain manual review rules. This level of granular control is a hallmark of open-source tools where the user can shape the tool to their specific workflow.

Moving forward, the author plans to clean up and share these custom extensions with the community. They encourage other developers who feel frustrated by the 'feature bloat' of major platforms to try open-source alternatives. The goal is to find a balance between powerful AI capabilities and a stable, predictable development environment that doesn't get in the way of the work.