What is Pi?
Pi is a terminal coding agent with a deliberately small core: a short system prompt, a handful of tools and no built-in sub-agents or plan mode. Everything else comes from extensions, skills, prompt templates and themes, which you can ask Pi to write for you and share as npm or git packages. It also runs in print and JSON mode, over RPC or as a TypeScript SDK.
Mario Zechner wrote it as pi-mono; in April 2026 he joined Earendil, the Vienna company founded by Armin Ronacher and Colin Daymond Hanna, and the project moved to earendil-works/pi under the MIT licence. The monorepo also holds pi-ai (one API across providers), pi-agent-core, pi-tui and Pi Durable. Pi 1.0 shipped on 1 October 2026.
Is Pi worth it?
Pi's argument is that a coding agent should be small and that you should add what you need. It ships without sub-agents, plan mode or a permission system, keeps its system prompt short to save tokens, and makes extensions first-class: a new command, tool, provider or TUI panel is a TypeScript file, often written by Pi itself mid-session. Sessions are stored as trees, so you can rewind to any earlier message, branch and share the whole thing as a gist.
Adoption is broad and not only on GitHub. The @earendil-works/pi-coding-agent package had 23,215,757 npm downloads in the 30 days to 13 November, a figure inflated by OpenClaw and other tools that embed Pi, and the repository has 122,011 stars on 7,970 commits. Pi 1.0 in October reversed Zechner's earlier position and added MCP support, along with a code-mode sandbox for tool calls and cache warming for Anthropic models. v1.5.2 shipped on 10 November.
The supply-chain work is better than most agents bother with. Direct dependencies are pinned, npm resolution refuses packages less than two days old, installs run with lifecycle scripts disabled and the pi.dev installer pins transitive dependencies too.
The catch: there is no sandbox by default. Pi runs with your user's full file, network and credential access, and the README says plainly to containerise it, with three documented patterns, if you want limits. New contributors' issues and pull requests are closed automatically and reviewed daily, which keeps the 221 open issues low but can feel unwelcoming.
Our take: Claude Code and Codex are appliances; Pi is a workbench. The time you spend building your own workflow pays back only if you will use it for months.
Where it earns its place
A senior developer who finds Claude Code's behaviour shifts between releases
Keep a coding agent whose prompt and tools stay the same until they change them
Install Pi, log in with a supported subscription through OAuth or an API key, and pin the version. Custom commands live in project extensions, so the workflow is versioned with the repository. No licence cost.
A startup building an agent feature into its own product
Reuse a tested agent loop and multi-provider model layer instead of writing one
Embed pi-agent-core and pi-ai through the TypeScript SDK or drive Pi over RPC, as OpenClaw does. MIT licensing means no fees; you pay only model usage.
A security-minded team lead at a fintech
Let developers use an agent without giving it the run of their laptops
Run Pi inside Docker or the Gondolin micro-VM extension so tools and shell commands are isolated while provider login stays on the host. Setup takes longer, but Pi itself has no built-in permission prompts.
Before you commit
- Setup
- `curl -fsSL https://pi.dev/install.sh | sh` (or npm with `--ignore-scripts`, Node.js 22.19+), run `pi` in a project and `/login` to connect a provider; minutes to the first task.
- Integrations
- Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, OpenRouter, Ollama and more; MCP since 1.0; skills and AGENTS.md; RPC, JSON mode and a TypeScript SDK; Nix flake.
- Data and privacy
- Code and prompts go to the provider you choose. Sessions are stored locally as JSONL files. Pi downloads model catalogue updates from pi.dev at runtime.
- Cost at scale
- No paid edition; cost is model usage or the subscription you log in with. The short system prompt keeps per-request overhead low.
- Limits
- Runs with full user permissions unless containerised; no sub-agents or plan mode out of the box; issues from new contributors are auto-closed.
What’s good
- Small core with first-class TypeScript extensions
- 15+ providers with API keys or subscription OAuth
- Tree-structured sessions you can branch and share
- Careful dependency pinning and script-free installs
What’s not
- No permission system or sandbox by default
- No sub-agents or plan mode without extensions
- First-time contributors are auto-closed
- Payoff depends on time spent customising
Best for
Developers who want a lean terminal agent they can extend, script or embed in their own product.
Pi alternatives
Pi: common questions
- Is Pi worth it?
- Pi is the coding agent for people who would rather shape their own tool than learn someone else's, and it has the engineering discipline to be trusted with that job. We rate it 4.3 out of 5.
- How much does Pi cost?
- Free and open source (MIT). You pay your model provider or log in with a supported subscription through OAuth.
- What are the downsides of Pi?
- No permission system or sandbox by default; no sub-agents or plan mode without extensions; first-time contributors are auto-closed; payoff depends on time spent customising.
- Who is Pi best for?
- Developers who want a lean terminal agent they can extend, script or embed in their own product.
- How much does Pi cost for a team or in production?
- No paid edition; cost is model usage or the subscription you log in with. The short system prompt keeps per-request overhead low.
- Is Pi safe for company data?
- Code and prompts go to the provider you choose. Sessions are stored locally as JSONL files. Pi downloads model catalogue updates from pi.dev at runtime.
- What are the best Pi alternatives?
- The closest alternatives we have reviewed are OpenCode, OpenAI Codex CLI and OpenClaw.
Answers checked Nov 14, 2026.
Make Pi? Show this rating on your site
