Pi
A minimal terminal coding harness you extend in TypeScript instead of configuring.
Pi gives the model four tools — read, write, edit, bash — and stops there. Sub agents and plan mode are deliberately absent; the project's answer is that you add what you need as a TypeScript extension, a skill, or a prompt template, and publish it as a Pi Package over npm or git. It ships as a monorepo whose pieces (agent runtime, unified multi-provider LLM API, terminal UI) are usable as libraries on their own, and it runs in four modes: interactive, print/JSON, RPC, and an embeddable SDK.
The question that sorts people here is not which agent is better but whether you are adopting one or building one. Pi ships four tools and stops; the runtime, the provider layer and the terminal UI are meant to be imported as libraries. If you read that as freedom, this is your tool. If you read it as homework, it is not. One thing to settle before shell access: it has no permission system by design, so the isolation has to come from you — a container, a micro-VM, something.
Our take
Worth it when you intend to build on the agent rather than only use it — the four packages are meant to be imported, and that is a genuinely different offer from a CLI with a plugin folder. The flip side is stated plainly in its own README: there is no built-in permission system, so it runs with whatever access the shell that launched it has, and the project's answer is to containerise. That makes it a poor fit anywhere the sandbox has to come from the tool.
Last checked 2026-09-08
Best for
- Teams embedding an agent loop into their own tooling rather than adopting a product
- People who would rather write fifty lines of TypeScript than search a settings page
- Workflows already running inside a container or micro-VM
Limitations
- No permission system; isolation is the operator's job, not the tool's
- No sub agents or plan mode out of the box — you add them or install a package
- Extending it means writing TypeScript, not editing config
When to pick something else
- Environments where the agent itself must enforce what it may not touch
- People who want a read-only review mode on day one without building it
- Teams with no appetite for maintaining their own extensions
Key features
- Four built-in tools; everything else is an extension you write
- Extensions, skills, prompt templates and themes packaged as Pi Packages
- Runs interactive, as print/JSON, over RPC, or embedded via SDK
- Session branching and compaction
- Subscription login (Claude Pro/Max, ChatGPT, Copilot) or API keys
Pi alternatives
If Pi is not the right fit, these solve similar problems.
- opencodeOpen-source terminal coding agent that runs on whichever model you point it at.Official: opencode.aiFree
- Claude CodeTerminal-based coding agent that runs commands and edits files directly.Official: claude.com/product/claude-codefrom $20/mo
- Codex CLIOpenAI's open-source terminal coding agent, rewritten in Rust.Official: developers.openai.com/codex/clifrom $20/mo
- ChatGPTThe most widely used assistant, with image generation and a large plugin ecosystem.Official: chatgpt.comfrom $20/mo
- ClaudeAssistant known for long-context reasoning and writing that needs less editing.Official: claude.aifrom $20/mo
- GeminiGoogle's assistant, tightly integrated with Search, Gmail and Workspace.Official: gemini.google.comfrom $20/mo