Hey HN,
I've been building with Claude Code daily and kept thinking: what if I could give my users the same experience — an AI agent that explores context, plans, executes tools, and builds things — but embedded natively in my own application?
So I built *Hive Agent*, an open-source TypeScript framework that reproduces the core Claude Code architecture as a library you can integrate into any app.
*What it does:*
- *Workspace system* — Give the agent any context via a virtual filesystem. Feed it your user's data, project files, API responses — whatever your app needs. The agent reads, writes, and searches through it using bash-like commands.
- *Built-in Explore & Plan agents* — Just like Claude Code, Hive auto-spawns sub-agents to investigate workspace data before acting and to design step-by-step plans for complex tasks.
- *Sub-agent orchestration* — Spawn specialized agents with different models (Claude, OpenAI), different tools, and structured I/O. A research agent can use GPT-4, while the main agent uses Claude.
- *Custom tools* — Define any tool your app needs. API calls, database queries, browser automation, document generation — the agent calls them as needed.
- *Stateless & serverless-ready* — No persistent state. Works in Firebase Functions, Vercel, AWS Lambda. Pass history in, get history out.
- *Execution tracing & cost tracking* — Full hierarchical traces with per-model token usage and cost breakdown.
- *Interactive mode* — Agent can pause and ask clarifying questions, just like Claude Code does.
*What you can build with it:*
- AI coding assistants scoped to your platform
- Document generators that understand your users' context
- Project scaffolding tools
- Support agents with access to your internal APIs
- Any workflow where an AI needs to explore data, plan, and act
*Quick example:*
```typescript
import { Hive, ClaudeProvider } from '@alexnetrebskii/hive-agent'
const agent = new Hive({
systemPrompt: 'You are a project generator...',
tools: [myApiTool, myDbTool],
llm: new ClaudeProvider({ apiKey: process.env.ANTHROPIC_API_KEY }),
workspace: new Workspace({
paths: { 'user/requirements.json': { value: userReqs } }
})
})
const result = await agent.run('Create a new project based on my requirements')
```
The agent will explore the workspace, plan the approach, create files, call your tools — all autonomously.
*Links:*
- GitHub: https://github.com/anetrebskii/hive-agent
- Install: `pnpm add @alexnetrebskii/hive-agent`
- MIT licensed
I'd love feedback from anyone building agent-powered features. What tools or patterns would you want built-in? Try it out and let me know what breaks.
