On November 8, a personal blog post hit the front page of Hacker News: "Goodbye Claude, Hello Pi." It picked up 643 points and 410 comments that day, and the count kept climbing past 800 in the days after. The post itself is dated September 21 — it sat quietly for seven weeks, then detonated.
I read the post, a good chunk of the 410-comment thread, and Pi's README. I have not switched to Pi myself, and this isn't a switch story. It's a look at what the switchers actually want, because the complaints in that post are more specific — and more checkable — than most breakup letters in tech.
The three complaints, in the author's own numbers
The author is Caleb, a developer at Salesforce by his own homepage, who had used Claude Code since it arrived in the Claude Pro subscription over a year ago. He liked it — he still says Anthropic built a genuinely cool product. Then he lists what wore him down, and every complaint comes with a count attached.
244 releases in a year, over three-quarters of them behaviour-changing. That's his count, not an independent audit — he says so in the post. But the shape of the complaint is what matters: each release could change the system prompt or tool descriptions, which "can substantially change how the LLM performs." The harness you configured in January is not the harness you're running in September, and nobody asked you.
A ~3,000-token Bash tool description. The default Bash tool's description runs to almost 3,000 tokens, most of it about using git and opening pull requests on GitHub. His point is simple: the models already know how to do this. The overhead buys little and costs context on every single request.
A bug open since December 2025, closed as a duplicate. He uses Claude Code inside VSCode, where you can connect the harness to the IDE and have highlighted text land in context automatically. Except only one session can be "connected" at a time. Open a second tab for another task and it can't receive text-selection events at all. Reported at least as far back as December 2025, closed as a duplicate of an unrelated bug. The official VSCode GUI extension allows multiple tabs, he admits — but half its built-in slash commands don't work in it, and subagent output layers into the main thread in chronological order.
Around those sits the friction he says epitomizes the whole decision. Claude Code ignored AGENTS.md, so users copied or symlinked CLAUDE.md files across hobby repos. A day after he made his decision, Anthropic announced Claude Code would finally start supporting AGENTS.md. He jokes they did it to get him to stay. Too little, too late.
What Pi actually is
Pi is a minimal, open-source terminal coding agent harness, MIT-licensed, created by Mario Zechner (the libGDX game framework author) and now maintained by Earendil. Its design thesis is that a thin harness around a capable model beats a thick one: four tools — read, write, edit, bash — and a system prompt of roughly 200 tokens. The rest is built by the user through TypeScript extensions, skills, prompt templates, themes, and lifecycle hooks.
It moves fast for a small project: 1.0 shipped October 1, 2026, 1.3.0 followed October 15, and 1.5.0 on November 2. Install is curl -fsSL https://pi.dev/install.sh | sh or an npm global install, and it needs Node.js 22.19 or newer. It talks to Anthropic, OpenAI, Google, DeepSeek, and more, plus OpenRouter and local models through Ollama, LM Studio, llama.cpp, or vLLM. You connect a subscription or API key with /login inside Pi.
The part the viral post barely touches
Pi deliberately does not have things Claude Code users take for granted, and the README says so in plain language.
No built-in permission system. The README: "Pi does not include a built-in permission system for restricting filesystem, process, network, or credential access. By default, it runs with the permissions of the user and process that launched it." There is no per-action approval prompt before it edits files. The README's own advice is to containerize or sandbox it. Community permission extensions exist, but they're add-ons, not the default. Switching on the assumption that Pi is safer out of the box gets it exactly backwards.
No sub-agents, no plan mode. The README says Pi "ships with powerful defaults but skips features like sub-agents and plan mode." If your workflow leans on those, you'd be rebuilding them with extensions or giving them up.
MCP works, despite what older comparisons say. Pi has a full MCP Servers guide and support arrived in 0.99.0 on September 29, before 1.0. Comparison pages claiming "no MCP" predate that version.
And the money question: only the harness is free. Claude Code runs on a $20–$200/month subscription; with Pi you pay for the model instead, a subscription connected via /login or your own per-token API keys. The honest saving is overhead — a ~200-token system prompt and four tool descriptions leave more of every request for your code. Whether your monthly bill drops depends on your token usage, not on the harness price.
What the thread itself said
The Hacker News thread is worth reading precisely because it isn't a victory lap. One commenter put it bluntly: the post has "a lot of why 'not cc' but I didn't see anything about why 'hello Pi'." Another pointed out that most of the complaints are about features Pi doesn't even support in the first place. Others who've daily-driven Pi for a month or more say they aren't looking back — and in the same breath warn that compaction fails about half the time.
That split is the real story. The switchers aren't really buying a feature list. They're buying stability and control: a harness that doesn't change under them 244 times a year, that they can extend in twenty minutes, and that stays out of the way. The price is giving up the guardrails — permission prompts, sub-agents, plan mode — and accepting that the safety boundary is now yours to build, in a container, before your first real task.
The author says it best himself, near the end of the post: "Later tools like Pi have the benefit of hindsight to see what works and what doesn't, what's a core feature and what can be cut." First movers pay the innovation tax. Second movers get to choose what to keep.
Sources: "Goodbye Claude, Hello Pi," caleb.software (Sep 21, 2026; Hacker News front page Nov 8, 2026, 643 points / 410 comments, rising past 800 after); Pi repository and documentation, earendil-works/pi; Hacker News discussion thread. The 244-release and ~3,000-token figures are the author's own counts, not an independent audit.
Top comments (0)