Earendil shipped Pi 1.0 — a hardened, minimal, extensible coding-agent harness — and, same day, an experimental package called Pi Durable for long-running agents that survive crashes and run across surfaces. Pi is already a daily driver for a large open community (Earendil cites hundreds of thousands of weekly users); 1.0 is the “dependable substrate” cut, not a greenfield rewrite. On Hacker News the Pi 1.0 post sat at #1 with roughly 825 points and 286 comments within hours of launch — the clearest builder-tool spike in the radar window after yesterday’s Gemini 4 Argon coverage.
The pitch is deliberate: keep the core small, wait until a pattern has proven itself, then adopt it. 1.0 adds Codemode (native MCP plus non-LLM models such as Jev and image models), virtual-model extensions, deferred tool loading, Anthropic cache warming, mid-conversation system messages, a new TUI theme, and full-screen mode by default. License: MIT. Install: curl -fsSL https://pi.dev/install.sh | sh (or npm @earendil-works/pi-coding-agent).
Image credit: Earendil
Key points
- What shipped: Pi 1.0 (stable coding-agent harness) + experimental Pi Durable (durable multi-surface agent framework). Both MIT-licensed.
- Who it’s for: Developers who want a malleable terminal coding agent they can extend — not a sealed vendor harness.
- What’s new in 1.0: Codemode (native MCP + non-LLM models); virtual models; deferred tool loading; Anthropic cache warming; mid-conversation system messages; new TUI theme; full-screen default.
- Scale signal (vendor): Hundreds of thousands of weekly users; GitHub repo earendil-works/pi at six-figure stars.
- Pi Durable (experimental): Crash-surviving tasks/checkpoints; memory / SQLite / JSONL storage; forkable concurrent conversations; pluggable extensions/tools/hooks; multiplayer watch/steer; ~15K LOC core (agent-readable).
- Install:
curl -fsSL https://pi.dev/install.sh | sh· npm@earendil-works/pi-coding-agent· Durable:npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord - Caveats: 9to6AI has not hands-on tested 1.0 or Durable. Pi Durable’s API may still change. Model access still requires your own provider keys/subscriptions. Community heat can outrun production readiness — verify on your stack.
What shipped
Earendil positions Pi as the harness you adapt to your workflow, not the other way around. Defaults are powerful; features other agents bake in (sub-agents, plan mode, permission gates) are meant to be built or installed as extensions.
| Field | Value |
|---|---|
| Product | Pi 1.0 (coding agent) + Pi Durable (experimental) |
| Org | Earendil |
| License | MIT |
| Docs | pi.dev |
| Repo | github.com/earendil-works/pi |
| Surfaces | Interactive TUI · print/JSON · RPC · SDK |
| Providers | 15+ (Anthropic, OpenAI, Google, Azure, Bedrock, xAI, OpenRouter, Ollama, …) |
| Pi Durable storage | Memory, SQLite, JSONL (+ conformance suite) |
Official announcements: Pi 1.0 · Pi Durable.
What changed vs prior Pi and the market
From “widely used” to “1.0.” The release letter is explicit: months of hardening from issue/PR feedback so people and businesses can depend on it. That is a maturity signal, not a new model drop.
Codemode + native MCP without abandoning minimalism. Earendil had argued publicly against bolting every protocol into the core. 1.0’s Codemode path — Pi writes scripts that orchestrate tools, with native MCP and support for decision/image models — is the compromise that “stuck.” Virtual models let you plan on one frontier model and implement on another (demoed with Jev deciding the switch).
Agent-harness competition is the story. Frontier labs keep pushing sealed TUIs (Claude Code, Codex, Antigravity/agy). Pi’s bet is the opposite: primitives, extensions, tree-structured shareable sessions, and mid-session model switching. HN’s jump on 1.0 fits a week where harness quality — permissions, compaction, steering — is as loud as model scores.
Pi Durable opens a second product shape. Coding Pi stays terminal-and-machine local. Durable adds storage-backed harnesses that resume after process death, run many conversations concurrently, fork threads, and let multiple clients watch and steer. That is closer to Slack bots, triage agents, and always-on workers than to a one-shot pi session — and Earendil marks it experimental on purpose.
What it means for builders
If you already live in Pi. 1.0 is the upgrade to take carefully: Codemode/MCP, cache warming for Anthropic, deferred tools, mid-conversation system messages. Re-check your extensions after /reload. Treat Durable as a separate experiment unless you need crash survival or multiplayer steering.
If you’re shopping harnesses after Claude Code / Codex / agy friction. Pi’s differentiator is malleability — ask Pi to build the feature, or install a package — plus open MIT code. Expect to bring your own model keys. Compare steering (Enter vs Alt+Enter), session trees, and extension APIs against your current tool before migrating a team.
If you need agents that outlive a laptop sleep. Pi Durable’s checkpointed tasks, safe tool replay, and SQLite/JSONL backends are the relevant surface. Start with the vacation-planner / coding-agent Durable examples; assume API churn.
If you care about decision models (Jev, Clef). 1.0’s Codemode path explicitly includes non-LLM models like Jev. That makes Pi a practical place to wire decision models into coding workflows — complementary to Cloudflare’s same-day Clef open-weight drop (runner-up on today’s radar).
If you need enterprise SSO / sealed compliance. Pi is an open harness, not a managed enterprise product. You own sandboxing, secrets, and policy. Durable’s ownership trees and hooks can encode approvals, but you build them.
What to watch
- Whether Codemode + MCP becomes the default path for Pi packages, or stays opt-in for power users.
- Pi Durable API stability and first production Slack/GitHub bot write-ups Earendil teased.
- How Pi’s extension ecosystem compares to Claude Code mods (also shipping in this window) for teams that want to rewrite harness behavior in TypeScript.
- Independent hands-on comparisons vs Claude Code / Codex / OpenCode on steering, compaction, and cost transparency (
/session).
