@0x6d61/wn-core

@0x6d61/wn-core

日本語

Lightweight AI Agent Core Framework for Node.js.

Build any AI agent by combining LLM providers, personas, skills, MCP tools, and built-in tools.

Features

• 4 LLM Providers — Claude, OpenAI, Ollama, and Gemini behind a unified interface
• Tool System — Built-in tools (read / write / shell / grep) + MCP dynamic tool loading
• 3-Layer Model — Persona (system prompt) / Skill (action definition) / Agent (sub-agent) hierarchy
• JSON-RPC 2.0 — Core and UI communicate over stdin/stdout as separate processes
• Worker Threads — Parallel sub-agent execution without blocking the main loop
• MCP Support — Dynamically load tools from Model Context Protocol servers

Install

npm install @0x6d61/wn-core

Global install (to use as CLI):

npm install -g @0x6d61/wn-core

Quick Start

CLI (JSON-RPC Server)

wn-core serve --provider claude --model claude-sonnet-4-20250514

Listens on stdin/stdout for JSON-RPC 2.0 messages from a TUI or any client.

# Test: pipe a JSON-RPC message
echo '{"jsonrpc":"2.0","id":1,"method":"input","params":{"text":"hello"}}' | wn-core serve

As a Library

import {
  createClaudeProvider,
  AgentLoop,
  ToolRegistry,
  createReadTool,
  createWriteTool,
  createShellTool,
  createGrepTool,
  createNoopHandler,
} from '@0x6d61/wn-core'

// 1. Create an LLM provider
const providerResult = createClaudeProvider(
  { apiKey: process.env['ANTHROPIC_API_KEY'] },
  'claude-sonnet-4-20250514',
)
if (!providerResult.ok) throw new Error(providerResult.error)

// 2. Register built-in tools
const tools = new ToolRegistry()
tools.register(createReadTool())
tools.register(createWriteTool())
tools.register(createShellTool())
tools.register(createGrepTool())

// 3. Build an AgentLoop
const loop = new AgentLoop({
  provider: providerResult.data,
  tools,
  handler: createNoopHandler(),
  systemMessage: 'You are a helpful assistant.',
})

// 4. Run a single conversation turn
const result = await loop.step('Read src/index.ts and summarize it')
if (result.ok) {
  console.log(result.data)
}

Configuration

Configuration is loaded from two levels: ~/.wn/config.json (global) and .wn/config.json (project-local). CLI flags take the highest priority.

{
  "defaultProvider": "claude",
  "defaultModel": "claude-sonnet-4-20250514",
  "defaultPersona": "default",
  "providers": {
    "claude": { "apiKey": "${ANTHROPIC_API_KEY}" },
    "openai": { "apiKey": "${OPENAI_API_KEY}" },
    "ollama": { "baseUrl": "http://localhost:11434" },
    "gemini": { "apiKey": "${GEMINI_API_KEY}" }
  },
  "mcp": {
    "servers": [
      {
        "name": "example",
        "command": "npx",
        "args": ["-y", "example-mcp-server"]
      }
    ]
  }
}

API keys can reference environment variables using ${ENV_VAR} syntax.

OAuth Token (authToken): Claude provider also supports OAuth authentication via authToken. Use authToken instead of apiKey when authenticating with a Claude OAuth token:

{
  "providers": {
    "claude": { "authToken": "${ANTHROPIC_AUTH_TOKEN}" }
  }
}

Priority (highest to lowest): CLI flags > project-local .wn/ > global ~/.wn/

Personas, Skills & Agents

wn-core uses a 3-layer resource model. Resources are loaded from ~/.wn/ (global) and .wn/ (project-local). Local resources override global ones with the same name.

Persona (System Prompt)

A persona is a plain Markdown file that defines the system prompt for the LLM.

Directory: personas/

~/.wn/personas/default.md
.wn/personas/security-expert.md

Example (personas/default.md):

You are a helpful assistant.
Answer concisely and accurately.

No frontmatter required — the entire file content becomes the system prompt. Select a persona by name (filename without .md) via defaultPersona in config or --persona CLI flag.

Skill (Action Definition)

A skill defines a reusable action with a description, allowed tools, and instructions.

Directory: skills/<skill-name>/SKILL.md

~/.wn/skills/code-review/SKILL.md
.wn/skills/recon/SKILL.md

Example (skills/code-review/SKILL.md):

---
name: code-review
description: Review source code for bugs and security issues
tools: [read, grep, shell]
---

## Instructions

1. Read the target files using the read tool.
2. Check for common vulnerabilities (injection, XSS, etc.).
3. Report findings with severity and remediation advice.

Frontmatter fields:

The body (after ---) contains the detailed instructions.

Agent (Sub-Agent)

An agent definition configures a sub-agent with its own persona, skills, provider, and model.

Directory: agents/

~/.wn/agents/scanner.md
.wn/agents/code-reviewer.md

Example (agents/scanner.md):

---
persona: security-expert
skills: [recon, port-scan]
provider: claude
model: claude-sonnet-4-20250514
---

A sub-agent specialized in network reconnaissance.
Runs port scans and service detection, then reports findings.

Frontmatter fields:

The body becomes the agent's description.

Architecture

wn-tui (separate process) <-- JSON-RPC 2.0 --> wn-core
                                                  |
                                                  +-- AgentLoop
                                                  +-- LLMProvider (Claude/OpenAI/Ollama/Gemini)
                                                  +-- Loader (persona/skill/agent)
                                                  +-- Tools (read/write/shell/grep + MCP)
                                                  +-- SubAgentRunner (Worker Threads)
                                                  +-- RPC Server (JSON-RPC 2.0)

See docs/architecture.md for details.

RPC Protocol

Core and clients communicate via JSON-RPC 2.0 over stdin/stdout (NDJSON format).

Core -> Client (Notification)

Client -> Core (Request)

Development

npm install
npm test            # Run tests (344 tests)
npm run typecheck   # TypeScript type check
npm run lint        # ESLint
npm run build       # Build (tsup)

Requirements

• Node.js >= 20

License

MIT