@svton/agent-core

Core runtime for Svton AI Agent. Provides a ReAct loop with tool execution, context management, multi-provider LLM support, MCP protocol, skills, memory, permissions, planning, subagents, and plugins.

Install

npm install @svton/agent-core

Quick Start

import {
  AgentRuntime, OpenAIProvider, ToolRegistry, setPlatform
} from '@svton/agent-core';
import { BrowserPlatform } from '@svton/agent-platform';

setPlatform(new BrowserPlatform());

const provider = new OpenAIProvider({
  baseUrl: 'https://api.openai.com',
  apiKey: 'sk-xxx',
  models: [{
    id: 'gpt-4o', name: 'GPT-4o', contextWindow: 128000,
    supportsToolUse: true, supportsVision: true, supportsStreaming: true,
  }],
});

const registry = new ToolRegistry();
registry.register(webSearchDef, new WebSearchExecutor());

const runtime = AgentRuntime.create({
  provider, model: 'gpt-4o', toolRegistry: registry,
});

const stream = runtime.run('Hello, what can you do?');
for await (const event of stream) {
  if (event.type === 'text_delta') process.stdout.write(event.text);
  if (event.type === 'done') console.log('\nDone');
}

Features

• ReAct Loop — Think → Act (tool call) → Observe → Think → ...
• Multi-Provider — OpenAI (and compatible: DeepSeek, Ollama, vLLM) + Anthropic
• Tool System — 20+ built-in tools (file, shell, web, memory, planning, computer use, Chrome CDP)
• MCP Protocol — Connect external tool servers via HTTP, SSE, or Stdio
• Skills — Discover and inject context-aware instructions
• Memory — Project-level (AGENT.md) + auto memory (IStorage)
• Permissions — 5 modes (read_only → auto) with rule engine
• Hooks — 8 lifecycle events (pre/post tool, session, etc.)
• Planning — Multi-step plans with dependency tracking
• Subagents — Spawn isolated agents with restricted toolsets
• Plugins — Install from directory or Git

Providers

OpenAI Compatible

new OpenAIProvider({
  baseUrl: 'https://api.openai.com',  // or DeepSeek, Ollama, etc.
  apiKey: 'sk-xxx',
  models: [/* ModelInfo[] */],
})

Anthropic

new AnthropicProvider({
  apiKey: 'sk-ant-xxx',
  // Default models: claude-sonnet-4, claude-haiku-4
})

Built-in Tools

| Category | Tools | Browser | |----------|-------|---------| | File | file_read, file_write, file_edit | No | | Search | grep, glob | No | | Shell | bash | No | | Web | web_search, web_fetch | Yes | | Memory | memory_save, memory_recall | Yes | | Planning | plan_create, plan_update_step, plan_get_status | Yes | | Computer Use | screenshot, mouse_click, mouse_move, keyboard_type, keyboard_press_key | Tauri only | | Chrome CDP | chrome_navigate, chrome_screenshot, chrome_click, chrome_type, chrome_evaluate, chrome_get_content | Browser only |

Agent Events

The run() method returns an AsyncGenerator<AgentEvent>:

• text_delta — LLM text output
• thinking_delta — Chain-of-thought (DeepSeek / Claude)
• tool_call_start / tool_call_progress / tool_call_end — Tool execution
• tool_approval_needed — Awaiting user approval
• context_compacted — Context window compression
• done — Run complete with token usage
• error — Runtime error

License

MIT