ToolScore

Pi

Earendil (Mario Zechner) · ai coding agents · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-10

Open-source minimalist CLI agent — 121k stars, model-agnostic, "the Emacs of coding agents".

#3 of 15 in AI Coding Agents

Try Pi for $0
Real price
$0 · OSS
open-source
BYOK (own key)
Yes
15+ providers (Anthropic, OpenAI, Google, Azure, Bedrock, Groq, xAI, OpenRouter, Ollama...) via API key or OAu
Limits
8.0/10
$0 software + your tokens (any provider)
Verdict
MAXIMUM FREEDOM
No limits of its own — your provider's rate limits are the only ceiling.

ToolScore breakdown — why 79.5/100

Price / Value
9.0
Token & Rate Limits
8.0
Speed / Latency
8.0
Privacy & Retention
7.0
BYOK (Own API Key)
9.5
Output Quality
7.5
Stability
8.0
Ease of Setup
6.0
Billing Transparency
9.5
Support & Community
7.0

10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →

What it does well

  • ▸121k GitHub stars and a huge HN following (608-point launch thread) — the most-customizable agent in the field
  • ▸15+ providers incl. Ollama and custom endpoints; swap models mid-session with /model
  • ▸Short system prompt makes it a great harness for small and local models
  • ▸100% MIT open source with documented supply-chain hardening

Where it costs you

  • ▸Deliberately spartan: no permission popups, no plan mode, no baked-in sub-agents — you assemble it yourself
  • ▸No built-in permission system — you must bring your own Docker/VM sandbox
  • ▸HN users report it "didn't finish the job" vs Claude Code on hard tasks

Hidden traps — the stuff the pricing page doesn't say

high riskNo permission system by default

Pi does not restrict filesystem, process, network, or credential access. Run it in Docker or a micro-VM before pointing it at sensitive repos.

medium riskBYOK means uncapped API bills

No subscription, no spend cap — an aggressive agent on Opus prices can burn money fast.

medium riskExtensions come from random repos

The pi install npm:... ecosystem has no vetting; core hardening is good but third-party skills are on you.

Try Pi

Know the limits before you pay: read the traps above, then decide with eyes open.

Visit official site

Affiliate link where available — never affects the score.

What the advertised price hides
Free MIT software, no paid tiers anywhere. You pay token prices at your provider.
Sources & freshness

Data as of 2026-11-10. Prices and limits change monthly — report anything stale. How we verify: methodology.

Closest alternatives

Pi review

In-depth analysis · last updated 2026-11-10

The harness you shape yourself

Pi is the open-source answer to the question "what if the agent were mine?" — a MIT-licensed CLI/TUI with 121k GitHub stars and an HN crowd calling it "the Emacs of coding agents". Everything is an extension: skills, themes, prompt templates, TypeScript add-ons installed with pi install npm:....

Model freedom

Fifteen-plus providers, OAuth or API key, local models via Ollama, and mid-session model switching (/model). Its deliberately short system prompt also makes it one of the best harnesses for small and local models — where Claude/Codex-style agents drown in their own prompt.

The honest downside

The minimalism is a choice, not an oversight: no permission system (bring Docker), no plan mode, no sub-agents out of the box — that's why the batteries-included fork oh-my-pi exists. And on hard tasks HN users report it "didn't finish the job" next to Claude Code; the harness is only as good as the model and the extensions you wire in.

Verdict

For developers who want total control and will pay for it with setup effort — 79.5/100. Run it sandboxed and pin your extension versions.

From the insights