edge-pi

edge-pi is a lightweight, Vercel AI SDK-based coding agent library and CLI. It provides the core primitives for building AI-powered coding assistants with tool support, session management, and context compaction.

What is edge-pi?

edge-pi consists of two packages that work together to provide a complete coding agent experience:

• edge-pi — The SDK. A TypeScript library built on the Vercel AI SDK that provides a CodingAgent class, file system tools, session persistence, and context compaction.
• edge-pi-cli — The CLI. An interactive terminal application (epi) that wraps the SDK with multi-provider support, authentication, skills, and both interactive and non-interactive modes.

Key Features

• Vercel AI SDK integration — Works with any Vercel AI LanguageModel (Anthropic, OpenAI, Google, and more)
• 7 built-in tools — read, write, edit, bash, grep, find, ls
• Session management — Tree-structured JSONL sessions with branching support
• Context compaction — Automatic summarization of old messages to stay within context limits
• Streaming and non-streaming — Both stream() and generate() execution modes
• Skills system — Extend the agent with custom instructions via Markdown files
• Multi-provider auth — API keys and OAuth for Anthropic, OpenAI, and Google

Quick Example

import { CodingAgent } from "edge-pi";
import { anthropic } from "@ai-sdk/anthropic";

const agent = new CodingAgent({
  model: anthropic("claude-sonnet-4-20250514"),
});

const result = await agent.generate({
  prompt: "Read the README.md and summarize it",
});

console.log(result.text);

See the Quick Start guide to get up and running.

Architecture

edge-pi is designed as a thin, composable layer on top of the Vercel AI SDK. The core CodingAgent class wraps the SDK's tool loop with sensible defaults for coding tasks:

CodingAgent
├── Vercel AI SDK (generateText / streamText)
├── Tool sets (coding, readonly, all)
├── System prompt builder
├── Message steering queue
├── Session persistence (JSONL)
└── Context compaction

Packages