Goodbye Claude, hello Pi
Many developers are moving from Anthropic’s Claude Code to Pi and similar “harnesses” that sit on top of LLMs, citing frustration with opaque behavior, rapid prompt/tool changes, vendor lock‑in, and subscription restrictions. Pi is praised for being minimal, open, and highly extensible—letting users customize tools, permissions, compaction, and even build full GUIs—though critics note it lacks built‑in safety rails and can become a maintenance project of its own. Across the thread, people weigh trade‑offs between TUIs and GUIs, open vs closed models, and native vs third‑party harnesses, with a recurring theme that control, security, and predictable behavior increasingly matter as much as raw model capability.
Pi vs. Claude Code (CC)
- Many commenters have migrated from Claude Code to Pi (or Oh My Pi), citing:
- CC’s frequent behavior changes due to constant system-prompt/tool tweaks.
- Opaque internals, hard-to-read or closed source, and difficulty understanding what it’s actually doing.
- Frustration with features like auto worktrees, subagents, and compaction that can get stuck or create messes.
- Pi is praised as minimal, transparent, and “yours”: easy to inspect, reason about, and extend; good session/tree management; strong telemetry about what agents are doing.
- Some still consider CC the best native harness for Anthropic models, citing strong integration, better compaction, and use of undocumented flags/headers that third‑party harnesses can’t match.
AGENTS.md, skills, and prompts
- The AGENTS.md vs CLAUDE.md issue is seen as emblematic: CC long ignored AGENTS.md, then added support but still prioritizes CLAUDE.md and lacks full support for
.agents/skills. - Several complain about extremely large system prompts in vendor harnesses (CC, Codex), which increase costs and reduce clarity; Pi’s smaller prompts are considered friendlier, especially to smaller/local models.
- Some argue rapidly changing prompts are “hill climbing” optimization; others see it as controlling users and undermining predictability.
Subscriptions, APIs, cost, and lock‑in
- Strong focus on Anthropic subscriptions:
- Subscriptions historically tied tightly to CC; third‑party harness use is constrained or ToS‑sensitive, with some fearing bans.
- Newer plans include API credits, but billing behavior (especially via
claude -por Agent SDK) is reported as confusing.
- Many use open or cheaper models (DeepSeek, GLM, Qwen, etc.) via Pi to drastically cut costs; others insist top closed models (especially Opus) still outperform open weights on hard tasks.
- Several note vendor lock‑in via subsidized tokens and native harness features (artifacts, remote compute, caching), but recommend tiered setups mixing cheap open models with frontier models only for hardest parts.
Security, sandboxing, and safety
- Pi is criticized for lacking built‑in “auto mode” / safety gates; defenders say this is by design and should be handled via extensions or OS‑level sandboxing.
- Many run agents inside Docker, Bubblewrap, containers, or restricted shells, often disabling permission prompts entirely to avoid fatigue.
- Various Pi extensions are cited to add permission systems, policy authorizers, auto‑mode–like behavior, and compaction strategies.
TUI vs. GUI and alternative harnesses
- Heated debate:
- TUI fans value SSH friendliness, low resources, dense info, easier sandboxing, and multi‑session workflows (often with tmux/herdr).
- GUI fans prefer better text rendering, artifacts, images, diffs, session organization, and mobile/desktop continuity.
- Many tools besides Pi and CC get discussed (DeepSeek Harness, OpenCode, Codex, Paseo, Orca, T3Code, custom GUIs), reflecting a fragmented but very active harness ecosystem.