$ cat ~/map/i-tried-pi-devto.md

I tried Pi after watching its founder explain why he quit Claude Code

A hands-on dev.to write-up of trying Pi after hearing its founder's case against Claude Code.

tutorials

Facts

repository: dev.to/urvvil/i-tried-pi-after-watching-its-founder-explain-why-he-quit-claude-code-2oef ↗
GitHub Stars: —
Forks: —
license: —
npm / mo: —
last push: —

## TL;DR

A hands-on dev.to article by Urvil Joshi: after watching pi creator Mario Zechner explain why he quit Claude Code, the author installs pi, walks its differentiating design choices, and delivers a comparative verdict — impressed, with one billing dealbreaker.

## Key points

$Framing: pi creator Mario Zechner's video argued that modern coding agents are bloated — per the article, Claude Code's system prompt has grown to roughly 14,000 tokens — while models are already trained to act as coding agents, so pi strips down to four tools and a tiny system prompt the agent can modify itself.
$Covers the on-ramp (one-line install, /login, /model, /settings) and three design choices the author found differentiating: the editable system prompt, tree-based sessions, and bash-only tooling — plus extensions (two demos) and skills.
$Concrete facts it records: pi's four default tools are read, write, edit, and bash, with no MCP servers, sub-agents, permission prompts, plan mode, or background bash by default; the system prompt is replaceable via system.md in ~/.pi/agent/ or appendable via append-system.md.
$Commands and layout it demonstrates include /tree and /fork for branching sessions, /reload for hot-reloading extensions, and /skill; extensions are TypeScript files in ~/.pi/agent/extensions/ and skills live in ~/.pi/agent/skills/.
$Verdict: the author values the minimalism, extensibility, and control over his own context, and calls it how he wants his coding agent to work — but his dealbreaker is that Anthropic bills pi sessions as extra per-token usage even on a subscription, so he keeps both tools side by side.
$Bottom line quoted in the article: if you want something that works on day one, Claude Code wins; if you want to actually own your workflow, pi wins.

## FAQ

Q Who is this article for?

Developers evaluating pi against Claude Code. It assumes no prior pi experience — it walks through install, login, first extensions, and skills before giving the comparison — and it is particularly useful for readers deciding whether a minimal harness fits their workflow.

Q What does the author ultimately conclude?

He stays impressed by pi's minimalism and control but keeps Claude Code alongside because of Anthropic's per-token billing of pi sessions on his subscription; he notes pi would be his full-time agent if he used ChatGPT, Copilot, or local models.

## References

## Related entries

Last updated: Aug 31, 2026