The Pi coding agent is Earendil's MIT-licensed, minimal coding-agent harness. Pi 1.0 shipped on October 1, 2026 as a stable daily-driver for developers who want a terminal agent they can shape, plus an experimental package called Pi Durable for long-running agent apps.
If you searched for Pi because you keep hearing Claude Code, Codex, or Cursor-class tools and want the open, minimal option, this page is the picker. You will see what 1.0 actually adds, when Pi Durable matters versus a one-shot terminal session, and how to install both from primary Earendil sources.
What Pi is (in one screen)
Pi is a local agent harness: the layer around a language model that reads files, runs commands, keeps a session, and stays extensible. Earendil's product site at pi.dev frames it as a harness you adapt to your workflow, not a sealed product that dictates one. Docs and install live there. Source is on GitHub under earendil-works/pi. Both Pi 1.0 and Pi Durable are MIT licensed, per Earendil's Pi 1.0 post.
The coding agent focuses on your machine (or a remote machine you drive), inside a terminal, with one person steering. If the process dies, you look at what happened and tell it to continue. That shape is intentional. Pi Durable is the separate experimental package when you need agents that survive crashes, run many conversations, and sit behind Slack or other surfaces.
What Pi 1.0 adds
Earendil's Pi 1.0 announcement (October 1, 2026) lists the features that graduated into the stable release after months of feedback:
- Codemode — native support for MCP, plus non-LLM models such as Jev and image models. Codemode is how Pi writes and runs scripts that boil tool work down, instead of stuffing every tool schema into the prompt forever.
- Deferred tool loading — tools load when needed so the context stays lighter.
- Extension support for virtual models — you can build routers that plan on one model and implement on another (Earendil's own demo routes with Claude Opus and GPT, with Jev deciding when to switch).
- Cache warming for Anthropic models — keeps prompt cache warmer across turns.
- Mid-conversation system messages — transcript-aware prompt and tool changes without restarting the session.
- Full-screen TUI by default — plus a new TUI theme.
Earendil is explicit that Pi stays minimal on purpose. Features that did not prove themselves stayed out. Sub-agents and plan mode are not baked into the core; you build them with extensions or install a package, as pi.dev documents.
Install Pi 1.0 from the announcement:
curl -fsSL https://pi.dev/install.sh | shOn Windows, Earendil documents a PowerShell one-liner from the same post. npm, pnpm, and bun global installs for @earendil-works/pi-coding-agent are also listed on pi.dev.
When Pi Durable matters (and when it does not)
Stay on the Pi coding agent when your job is: open a repo, steer one session, ship the change, close the terminal. That is the product Earendil hardened into 1.0.
Reach for Pi Durable when you need durability beyond one process. Earendil's Pi Durable post describes it as a framework for long-running, malleable agent apps that can run anywhere a JavaScript runtime exists. It shares principles and packages with the coding agent (including pi-ai), but it does not replace the terminal agent.
Concrete Durable capabilities from that post:
- Storage backends (memory, SQLite, JSONL) with checkpoints so a dead process resumes unfinished tasks
- Many concurrent conversations, including forks (for example a Slack channel plus a thread)
- Extensions that bundle tools, hooks, system prompt sections, and durable tasks
- Background compaction for long transcripts
- Multiplayer views so several clients watch and steer the same conversation
Install the experimental package as Earendil documents:
npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chordAPI still may change. Treat Durable as experimental until Earendil marks otherwise.
Picker: Pi vs Claude Code, Codex, and Cursor-class tools
Use this as a fit check, not a scorecard of invented benchmarks.
Choose Pi when you want an MIT open harness, provider-agnostic models (pi.dev lists 15+ providers), a terminal-first daily driver, and the right to rewrite the agent with extensions. Codemode/MCP and deferred tools are the 1.0 reasons to revisit if you tried Pi earlier.
Choose Claude Code when you want Anthropic's proprietary coding agent surface and you are fine living inside that product boundary.
Choose Codex-class / ChatGPT coding surfaces when your team already centers OpenAI's tooling and you need that vendor's integration path more than a minimal open harness.
Choose Cursor-class IDE agents when your primary loop is editor-native autocomplete and chat inside the IDE, not a terminal harness you own.
Pi is strongest as the substrate you customize. Claude Code, Codex, and Cursor-class tools are strongest as polished product loops. Many developers keep more than one.
For a wider map of coding agents (without treating this page as a hub rewrite), see our roundup of the best AI coding agents.
Soft next step for developers
If you already write production code and you want structured practice on agents, RAG, and production workflows — not another sealed chat box — explore AI Flex. Self-paced paths, cancel anytime, built for developers adding real AI systems. Use the form below if you want the details.
Go deeper
Sources (page-verified)
- Earendil, Pi 1.0 (October 1, 2026)
- Earendil, Pi Durable (October 1, 2026)
- Docs and install: pi.dev