Introduction

Build intelligent AI apps with the composability of modern software development.

The Problem

AI Development Is Still Alchemy, Not Engineering.

Every AI developer knows these pain points:

“It worked in my notebook”
Your agent performs flawlessly in isolation, breaks immediately in production.

“The context window exploded… again”
Mid-execution crashes because someone added one more tool to the chain.

“Latest model broke everything”
Model updates shouldn’t require rewriting your entire application.

“Let me copy that agent from last project… somewhere”
No package manager, no versioning, no reusability. Just scattered prompts and prayers.

“Is it safe? Who knows?”
Security by obscurity isn’t security. But that’s all we have with prompt chains.

The Evolution of AI Development

We’ve watched AI development evolve through distinct generations:

1st Generation: Prompt Engineering
Just write better prompts — Artisanal craftsmanship that doesn’t scale. Every developer reinventing the wheel.

2nd Generation: Orchestration Frameworks
Chain all the things — Created more problems than they solved. The return of spaghetti code.

3rd Generation: Expert Stack
Build AI like we build software — Real engineering. Composable, versionable, deterministic.

Why Perstack?

Perstack introduces the Expert Stack, a software architecture that creates structured relationships between AI agents and applications, turning them into first-class software components, with reproducibility, observability, and composability.

Here’s what sets it apart:

💻 Built for App Developers
Build locally or in the cloud. Use the CLI or SDK. Frictionless setup.
🧠 Declarative by Design
Define experts in natural language. No tangled orchestration logic.
📦 Reuse, Like Code
Experts are modular, versioned, and publishable — just like npm packages. Compose them freely.
🧩 Composable, Scalable Experts
Delegate tasks across Experts. Architect deep reasoning flows without bloated contexts.
🔍 Observable and Reproducible
Every step, decision, and tool call is checkpointed. Resume, inspect, or rerun with confidence.
🔒 Security is the #1 Priority
Experts run in a sandboxed environment with limited system access by default.
🤝 Model Agnostic, Open by Default
Use OpenAI, Claude, Gemini, or local LLMs. Bring your own tools and infrastructure.

Learn more about the Expert Stack.

Quick Start

Ask the community’s @perstack/deep-research expert about how Perstack is different from other frameworks.


export ANTHROPIC_API_KEY="..."
export EXA_API_KEY="..."
 
npx perstack run @perstack/deep-research \
  "Research Perstack and other AI frameworks, compare them, and write a report on workspace."
 
# What just happened:
# 1. Expert connected to Exa via MCP
# 2. Started gathering information on search engines
# 3. Delegated to @perstack/web-search to gather more information
# 4. Created a report in workspace
# 5. Every step was checkpointed and reproducible

Alternatively, you can use the Perstack SDK to build your own experts.


npm install @perstack/runtime


pnpm add @perstack/runtime


yarn add @perstack/runtime


bun add @perstack/runtime


import { run } from "@perstack/runtime";
 
const result = await run({
  setting: {
    expertKey: "@perstack/deep-research",
    input: { text: "Research Perstack and other AI frameworks, compare them, and write a report on workspace." },
  },
})

Core Principles

Packaged Experts
Every Expert in Perstack is a versioned, shareable package—just like npm modules. Import community experts, or build your own.
Expert Composition
Experts delegate tasks to other Experts, creating sophisticated multi-expert systems without complexity. Each Expert maintains its own context window and specialized knowledge.
Deterministic Runtime
Every expert run produces checkpoints: serialized snapshots of execution state. Resume interrupted work, fork from any point, or replay entire workflows.
Observable by Design
Every step, tool call, and decision is logged and traceable. Debug AI workflows like traditional software — with full visibility into the reasoning process.
Runtime Security
Experts run in a sandboxed environment with limited system access.

What You Can Build

Agentic RAG with Your Own Data
Build NotebookLM-style experts that work with your docs, code, or knowledge base—with reasoning, not just retrieval.
Coding Experts by Your Own Conventions
Create experts that write code your way: follow your style guide, use your libraries, apply your patterns.
Enterprise-Grade AI Workflows
Multi-expert systems for complex tasks: process data, ensure compliance, generate reports—all automated.
Custom AI Tools
Package domain expertise as shareable experts: legal analysis, medical research, financial modeling.

And anything else you can imagine. If you can describe it, you can build an Expert for it.