Quick Answer: A Claude Code alternative worth testing is Pi, an open source coding harness that one longtime user adopted in September 2026. The switch came after accumulating frustrations: a multi-tab IDE limitation, a steady stream of system prompt changes, and feature sprawl. Pi offers a smaller footprint, more stability, and deep customization through extensions. The trade-off is setup effort and giving up Anthropic's official tooling.
Last year, one developer ran 244 releases of Claude Code through their own tally and found that over three-quarters of them changed the system prompt or tool descriptions. That's the kind of churn that quietly reshapes how your agent behaves from one week to the next. In a post shared on Hacker News, the author explains why they retired Claude Code as their main harness and moved their development setup to Pi. Their reasons aren't dramatic. They're the accumulation of small frustrations that add up.
The AGENTS.md announcement came too late
The timing is almost funny. The author decided to leave, and a day later Anthropic announced that Claude Code would support AGENTS.md files. Did that fix the problem? For this author, no. The issue was never a single missing file format. It was a pattern of friction that made the tool feel like it was working against their setup.
Here's the friction in practice. Many developers have built up CLAUDE.md files across their repositories, and the AGENTS.md convention has been spreading as a shared format for agent instructions. If you've invested in one, moving to the other means copying content or maintaining symlinks between the two. Neither option is hard, but both are maintenance work you didn't ask for.
The lesson here is that a late feature rarely repairs trust. Once a user has spent months routing around a tool's rough edges, each new release gets judged against that history.
The IDE session limit breaks review workflows
This is the part that matters most for anyone who reviews code before committing. The author uses Claude Code inside VS Code, where the harness connects to the editor. You highlight text, and it lands in the model's context automatically. That's excellent for review, and it's exactly why the limitation hurts.
Only one session can be connected to the IDE at a time. Open a second tab to start another task while the first is running, and the new session doesn't receive text selection events from the editor at all. The author reports this was filed at least as far back as December 2025 and closed as a duplicate of an unrelated bug.
The official Claude Code VS Code GUI extension supports multiple tabs, with selections loaded into each. But the author lists several gaps there too. Half of the built-in slash commands don't work in the GUI. There's no custom status line support. Subagent output is layered into the main chat in chronological order, which becomes hard to follow once more than one subagent runs at once. Here's where it gets tricky: each workaround solves one problem and exposes another.
More features, less clarity about what matters
Dynamic workflows, agent teams, and ultrareviews sound impressive in a release note. The author's objection is that they don't address what they needed, and they seem to push toward heavier token consumption. That's a fair question to ask of any vendor roadmap.
This is where the popular advice goes wrong. The common assumption is that a fast-moving tool with lots of new features is a healthy sign. For a power user who values stability, the opposite can be true. Every new feature is a new surface for behavior to change, and every change to the system prompt or tool descriptions can alter how the model acts on instructions you've already tuned.
The author's broader critique is about product direction. They describe a product where engineers and PMs add whatever suits their own workflows, with no clear owner whose job is to say no. You don't have to agree with that diagnosis to see the risk. Products without a strong "no" function tend to accumulate features faster than they accumulate coherence.
System prompt churn and token overhead
This is where the argument gets concrete. The author points to a system prompt cut by 80% in a separate analysis, yet still finds that Claude Code carries far more overhead than Pi. The default Bash tool description alone is almost 3,000 tokens, and most of it covers git usage and opening pull requests on GitHub.
The author's argument is that current models already know how to do these things without detailed instructions. That claim is plausible for common workflows, but it isn't universal. Some teams have strict commit conventions or safety rules that deserve explicit instructions. The honest test is to measure your own tasks before and after trimming.
The cost angle is real but depends on your scale. The author runs a back-of-envelope calculation: at 1 million daily active users and 100 million requests per day, trimming 1,000 tokens per request could save millions of dollars a year in inference costs. That's a hypothetical, and your actual savings depend on pricing, caching, and traffic. Still, it shows why token overhead is more than a cosmetic concern.
Here's a quick checklist for evaluating overhead in any harness:
- Measure the tokens consumed by the system prompt and tool descriptions before your first prompt.
- Log which tool descriptions are actually invoked in a typical session.
- Compare task success rates after removing instructions the model already handles well.
- Watch for behavior changes after each harness update, not just feature additions.
Claude Code vs Pi: when to stay and when to switch
The right choice depends on what you value. Here's how the author's experience compares across the factors they highlighted.
| Factor | Claude Code (per the author) | Pi (per the author) |
|---|---|---|
| Multi-tab IDE support | One connected session; GUI extension supports tabs but lacks some slash commands | Author built a multi-tab selection extension in about 20 minutes |
| Release churn | 244 releases this year; over 75% changed system prompt or tool descriptions | Described as stable, with a minimalist approach |
| Customization | Limited to supported configuration | Extensible; author has written several extensions for their own use cases |
| Vendor backing | Official Anthropic product tied to Claude Pro access | Open source; no vendor-backed roadmap |
Stay with Claude Code if you rely on vendor support, need features like agent teams, or your team's workflows already depend on its behavior. Switch if you're hitting the same IDE limits, or if you value a stable harness you can reshape. Note that the author's first few days with Pi "required some finagling," so budget time for the transition.
How to migrate your setup without losing a week
Migration is mostly about inventory and reversibility. Don't burn your bridges on day one.
- Inventory your instruction files. List every CLAUDE.md and any custom settings across your repositories.
- Decide on a single source of truth. Either maintain AGENTS.md as the canonical file or keep CLAUDE.md and link the two. Pick one and document it.
- Pilot on one repository. Choose a side project with low stakes and run your normal review loop.
- Write the missing pieces as extensions. If a feature you rely on is absent, the author's experience suggests a small extension can close the gap quickly.
- Keep the old tool installed for a week. You'll find edge cases you didn't plan for, and having a fallback removes the pressure to rush.
Here's where most guides stop, and it's the part that saves the most pain: compare outputs on the same prompt in both tools. Differences in review quality matter more than differences in feature lists.
The caveats worth keeping in mind
This is one developer's experience, and the author says so. The release count is their own tally, the token figures come from their analysis, and the cost math is a hypothetical. None of that means Claude Code is a bad product. The author says plainly that Anthropic has built something cool, and that building an early coding harness is hard.
There's also a real trade-off. Switching to an open source harness means you own more of the upkeep, and you lose the assurance that comes with an official product. For a solo developer who enjoys tinkering, that's a feature. For a team with compliance requirements, it may be a dealbreaker.
Frequently Asked Questions
what is a claude code alternative worth trying?
Pi is one open source option that the author in this post adopted after weighing stability, IDE workflow, and token overhead. It's worth trying if you want a minimal, customizable harness and can accept more setup work. Test it on a low-stakes repository first, and compare review quality against your current tool before deciding.
how do I switch from claude code to pi?
Start by inventorying your CLAUDE.md files and settings, then choose AGENTS.md or a linked format as your instruction source. Pilot on one repository, write extensions for any missing features, and keep Claude Code installed for about a week as a fallback. Compare outputs on identical prompts before fully migrating.
why do developers leave claude code?
Common reasons include IDE limitations such as single-session editor connections, frequent changes to system prompts and tool descriptions, and feature additions that don't match a user's workflow. Token overhead also matters to some teams. The author in this post cited all of these, with the IDE session limit and prompt churn as the most consequential.
can claude code read agents.md files?
According to the post, Anthropic announced AGENTS.md support for Claude Code one day after the author decided to leave. Check current Claude Code documentation to confirm how it handles AGENTS.md alongside CLAUDE.md in your version. Until you verify, keep your instruction files in a format you can maintain across both tools.
Switching harnesses is a bigger decision than picking a new editor theme, so treat it as an experiment. Try Pi on a side project this week, run the same prompt through both tools, and note where the review quality differs. Pass this to anyone wrestling with a bloated coding agent who's wondering whether a Claude Code alternative is worth the effort.