Longtime Claude Code user switches to open source Pi harness

A personal blog post on caleb.software, titled "Goodbye Claude, Hello Pi", explains why its author retired Claude Code as a main coding harness. The author has used the tool since it became available in the Claude Pro subscription plan, more than a year ago, and still calls it a really cool piece of software that will probably be remembered as a major driver of the current era of software development. But an accumulation of small frustrations led to the decision to move to Pi, an open source harness.

The trigger, in the author's telling, was AGENTS.md. Anthropic announced that Claude Code would start supporting AGENTS.md files a day after the author made the decision; the author jokes that this was meant to keep them, and says it came too little, too late. The author sees the episode as typical: having CLAUDE.md files scattered around repos is good guerrilla marketing, but users have to copy or symlink between the two formats.

The biggest practical complaint concerns VSCode. The author prefers to review most generated code before committing, and uses the IDE connection to highlight text in a file and have it included in context. But only one Claude Code session can be connected to the IDE at a time; a second tab opened for another task cannot receive the text-selection events. The author says this was reported at least as far back as December 2025 and closed as a duplicate of a totally unrelated bug. The official VSCode GUI extension does allow multiple tabs with selection loaded into each, but has its own problems: half of the built-in slash commands do not work in it, it has no support for a custom status line, and subagent output is layered into the main chat thread in chronological order, which is hard to read when more than one subagent runs.

More broadly, the author thinks Anthropic has worked at a feverish pace to add features, such as dynamic workflows, agent teams and ultrareviews, without much care for quality or stability. Those features, the author says, are mostly ways for enterprise customers to burn more tokens, and not something the author wants or needs.

The second theme is churn in the prompts. By the author's count, there have been 244 Claude Code releases this year as of writing, and over three quarters of them changed the system prompt or tool descriptions. The author says such changes can substantially alter model behaviour, especially when the new instructions contradict user prompts or repo conventions. Despite chopping 80% of the system prompt, Claude Code still carries more overhead than Pi, the author says, for seemingly no benefit. The default Bash tool description is almost 3,000 tokens, most of it about using git and opening GitHub pull requests, which the author argues the models already know how to do without detailed instructions. A footnote adds that the description tells the model to use cat <<'EOF', which before the widespread rollout of auto mode would always trip manual review rules regardless of existing git command allowlists. Another footnote offers a hypothetical: assuming 1 million daily active users and 100M requests per day, trimming 1,000 tokens from that description could save millions of dollars a year in inference costs.

The author's overall verdict is that Claude Code feels like a product without a strong vision or direction, as if excited engineers and product managers tack on features that suit their own workflows. It seems to the author that nobody high up is tasked with saying no to such requests. The author also apologises for the tone, says it is not their usual style, and says they are not under any delusion that they would have run the product better, since LLM coding harnesses are a new field and later tools like Pi benefit from hindsight.

On Pi, the author says the first few days needed some finagling, but it has been smooth sailing since. They enjoy its stability, minimalism and extreme customizability. After installing it, they wrote an extension that handles the multi-tab IDE functionality in about 20 minutes, plus a couple of other extensions for their own use cases, which they hope to share once cleaned up. The post ends by encouraging readers with similar frustrations to try an open source harness.

Key facts

The author, a Claude Code user since it joined the Claude Pro plan over a year ago, has retired it as the main coding harness and moved to the open source harness Pi.
Only one Claude Code session can connect to VSCode at a time; the author says the bug was reported at least as far back as December 2025 and closed as a duplicate of an unrelated one.
By the author's count, Claude Code had 244 releases this year, and over 3/4 of them changed the system prompt or tool descriptions.
The default Bash tool description is almost 3,000 tokens; the author says Claude Code still has more overhead than Pi despite cutting 80% of the system prompt.
The author says Pi had a rough first few days but has been smooth since, and an extension for multi-tab IDE support took about 20 minutes to write.

Why it matters

This is one heavy user's account, not a new finding, but it names concrete friction points in a widely used coding agent: a single-session IDE connection, a GUI extension with broken slash commands, and a system prompt and tool descriptions that change in most releases. It also shows a pattern the author calls out directly: users who want stability, a small prompt and the ability to extend the tool are looking at open source harnesses such as Pi.

Who it affects

Developers who run Claude Code inside VSCode and want several sessions with IDE text selection at once, and anyone who keeps personal prompts, repo conventions or allowlists that a prompt change could override. It also touches teams that keep both CLAUDE.md and AGENTS.md files, who currently have to copy or symlink between formats. Anthropic's announcement of AGENTS.md support, as described by the author, addresses part of that last complaint.

How to use it

The post is an opinion piece, not a tutorial. The author says Pi is open source, took a few days of finagling to settle into, and is highly customizable through extensions. The author's own Pi extensions, including one for multi-tab IDE support, are not yet released; they say they hope to share them after cleaning them up. Readers with the same frustrations are encouraged to try an open source harness.

How solid is it

It is a single personal blog post. The figures of 244 releases and over three quarters touching the system prompt or tool descriptions are the author's own count, with no method given. The 1 million daily users and 100M requests per day behind the cost estimate are an explicit hypothetical, not Anthropic data. The post offers no benchmark comparing Pi and Claude Code, and it does not say which other harnesses were considered. The IDE bug is described but its issue is not linked.

Risks and caveats

Much of the argument is opinion: that Anthropic packs in features without care for quality, and that nobody is tasked with saying no. The author hedges these with phrases like it seems like and feels like, and admits the post is a lot of hating and that they would not necessarily have done better. No response from Anthropic is given. The post says Pi has been smooth, but it describes a rough start and offers no evidence beyond the author's own experience.

“Ultimately, CC feels like a product without a strong vision or direction.”

— Author of the caleb.software post "Goodbye Claude, Hello Pi"