PI from Scratch

You get a fantastic, hands-on tutorial for building a simple coding agent in TypeScript, walking you through the core concepts of agent loops, tool use, and state management. The project is explicitly for learning; it is a 600-line teaching example, not a production-ready tool designed to solve complex coding tasks on a real repository. It lacks sandboxing, so running its run_bash command on your machine carries risk.

Pick this if you want to learn how to build your own agent from first principles. If you need an agent to do actual work for you today, pick a production tool like Aider or Cursor instead.

The agent executes shell commands. The architecture lacks a sandbox. This means a hallucinated or malicious command runs with the user's full permissions, creating a significant security risk. The repository documents run_bash as a core tool. Its purpose is educational, designed to teach agent mechanics by stripping away production safeguards. It is not intended for unattended use on any important project.

It succeeds as a tutorial. The project provides a minimal, 600-line TypeScript implementation that walks a developer through the core loop of an agent. It is a clear explanation of agent fundamentals.

PI from Scratch is documented as a tutorial and a minimal implementation inspired by the 'pi' agent. Its architecture is explicitly deconstructed into five modules: 1. llm.ts for model communication, 2. agent.ts for the core loop, 3. tools.ts for stateless functions, 4. tui.ts for terminal interface, and 5. cli.ts for orchestration. The separation of concerns between the agent loop and the presentation layer is a principled design choice.

As an educational resource, its value is in its legibility. It is not intended for practical application, a fact reflected in its lack of sandboxing, git operations, or formal benchmarks. Its purpose is to be understood, not to perform.

This is a classic developer-as-educator project. It's a sharp, well-structured tutorial for building a coding agent, essentially a deconstructed and simplified take on the 'pi' agent. The value here is entirely in the pedagogy; it's designed to be understood and replicated, which by definition means there's no moat. The GitHub stars and API credit sponsorships from OpenModel and Cubence are positive signals of community adoption and usefulness, but they don't point to a business model.

The trajectory for a project like this is well-established. It serves as an excellent portfolio piece for the creator, 'SaladDay', who will likely be hired by a developer tools company or an AI lab. The project itself will remain a static educational artifact, its maintenance dependent on the creator's spare time. It's not a venture, it's a very good resume.

This is an open-source tutorial for building a coding agent, not a deployable tool. It consists of five TypeScript files and instructional text intended to teach agent architecture. There are no enterprise features, no support contract, and no vendor entity to engage for procurement.

As educational material, it carries instructional value. As a production system, it offers zero reliability, no support, and requires us to own the entire stack, including the LLM endpoint and any security liabilities from local command execution. Not yet.

PI from Scratch is exactly what it says: a tutorial. It's a 600-line TypeScript project that walks you through building a basic coding agent. The code is split into five files (llm.ts, agent.ts, etc.) which I appreciate. You bring your own OpenAI-compatible key, so you control the model and cost. It's all MIT-licensed, so you can read it, fork it, and build on it without asking permission. It's a teaching tool, not a daily driver.

This isn't a framework you'd use for a real project; it's the source code for one. It lacks sandboxing, so running run_bash is a risk you accept. But for understanding the core agent loop—ask model, get tool call, execute tool, report result, repeat—it's solid. It's the blueprint you'd study before trying to build your own thing, and that's more valuable than most shrink-wrapped agent frameworks.