Introduction

kenpachi is a small TypeScript SDK for building tool-using LLM agents — without pulling in a heavy vendor framework. You get a typed agent loop, pluggable providers (Anthropic & OpenAI over raw fetch), and a handful of production features that most minimal agent loops skip.

Start here if you’re evaluating: read this page, then follow Quickstart. You’ll have a working agent with tools in under 5 minutes.

What you get out of the box

Feature	What it solves
Typed tools	Define tools once with Zod — the model gets a JSON schema, you get runtime validation.
Argument pre-coercion	Primitive argument types (like stringified numbers and booleans) are automatically coerced before schema validation keeps your agent loop resilient.
Time-travel context	Snapshot every turn; branch or rewind without re-calling the model for history you already paid for.
Saga rollback	Register undo handlers per tool — if a later step in the same batch fails, earlier steps are reversed.
Streaming	Token-by-token output via `agent.stream()` or a one-line `onText` callback on `run()`.
Handoffs	Wrap specialist agents as tools — a triage agent delegates to billing or support automatically.
Dynamic tool synthesis	Let the model author small sandboxed helpers; secrets stay in your connector registry, never in model output.

Minimal example

This is the whole mental model: provider → tools → agent → run.

import { z } from "zod";
import { Agent, defineTool, createAnthropicProvider } from "kenpachi";

const getWeather = defineTool({
  name: "get_weather",
  description: "Get current weather for a city",
  schema: z.object({ city: z.string() }),
  async execute({ city }) {
    return { city, tempC: 24, condition: "sunny" };
  },
});

const provider = createAnthropicProvider({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  model: "claude-sonnet-4-6",
});

const agent = new Agent(provider, [getWeather]);
const result = await agent.run("What's the weather in Nashik?");

console.log(result.text);
// → "It's sunny and 24°C in Nashik."

No boilerplate loop, no manual tool-call parsing — agent.run() handles the back-and-forth until the model finishes.

How it compares

Most teams building agents today either:

Roll their own loop — fast to start, painful when you need undo, argument coercion, streaming, or multi-agent routing.
Adopt a large framework — powerful, but heavy and opinionated.

kenpachi sits in the middle: small surface area, production-shaped defaults. You keep control of your stack (plain TypeScript, zero vendor SDK deps) while getting the guardrails that matter once you ship to real users.

When kenpachi is a good fit

You’re building a chat product and need undo / edit / regenerate without duplicate API costs.
Your agents call external APIs via tools and you want validation + rollback when things go wrong.
You want specialist sub-agents (billing, support, research) behind a single front-door agent.
You need streaming text in your UI without rewriting your agent loop.

Next steps

Install

Add kenpachi to your project and set up API keys.

Quickstart

Build your first tool-using agent in a few lines.

Streaming

Stream tokens to your UI as the model responds.

Handoffs

Route work to specialist agents automatically.

Getting Started

Core Concepts

API Reference

What you get out of the box

Minimal example

How it compares

When kenpachi is a good fit

Next steps

Install

Quickstart

Streaming

Handoffs

​What you get out of the box

​Minimal example

​How it compares

​When kenpachi is a good fit

​Next steps

Install

Quickstart

Streaming

Handoffs

What you get out of the box

Minimal example

How it compares

When kenpachi is a good fit

Next steps