Agent Smith

A toolkit to create local-first human-friendly agents in the browser or terminal

Feature	Description
🧠 Think	LLM inference via any OpenAI-compatible backend, first class Llama.cpp support
🔧 Work	Execute tools recursively — actions, workflows, commands, agents
💾 Remember	Semantic memory (LanceDB) + transient memory (IndexedDB/localForage)
💻 Interact	Terminal REPL (lm), WebSocket client, or Koa server API

Philosophy

• Composable — limited responsibilities, packages work together
• Declarative — focus on business logic, express features simply
• Explicit — simple and under user control, no hidden magic

Documentation

For AI Agents

• Codebase Summary — Architecture, key files, and patterns
• Documentation map — Complete documentation and usage guide

For Humans

📚 Documentation | 🧩 Plugins | 🎨 UI

Quickstart

Get Started

🖥️ Web UI

💻 Terminal Client

Packages

Package	Version	Description	Node.js	Browser
@agent-smith/cli		Terminal REPL client (lm command)	✅	❌
@agent-smith/agent		Agent runtime with inference loop	✅	✅
@agent-smith/core		Runtime engine (SQLite, config, tools, MCP)	✅	✅
@agent-smith/types		Shared interfaces and types	✅	✅
@agent-smith/smem		Semantic memory (LanceDB + embeddings)	✅	❌
@agent-smith/tmem		Transient key-value store (localForage/IndexedDB)	✅	✅
@agent-smith/wscli		WebSocket client with auto-reconnect	✅	✅
@agent-smith/server		Koa backend (REST API + WebSocket)	✅	❌

Plugins

Plugins extend the CLI with additional capabilities. Browse all plugins at agent-smith-plugins.

Category	Plugin	Package	Description
💻 Code Management	git	@agent-smith/feat-git	AI-powered commit messages, diff analysis
	sqlite	@agent-smith/feat-sqlite	Database operations, schema extraction
⚙️ System	fs	@agent-smith/feat-fs	Filesystem read/write with path authorization
	shell	@agent-smith/feat-shell	Sandboxed Docker execution (shell + Python)
🌐 Web	search	@agent-smith/feat-search	Multi-backend web search (DuckDuckGo, Wikipedia, crawl4ai)
	video	@agent-smith/feat-video	YouTube transcript extraction and chat
📝 Docs	autodoc	@agent-smith/autodoc	AI-powered documentation Q&A

Install a plugin: npm i -g @agent-smith/feat-shell, then add it to your config.yml.

Architecture Highlights

• SQLite-driven runtime: All configuration (backends, features, settings) lives in SQLite — zero hardcoded defaults
• Feature discovery: Agents, actions, workflows, and commands are YAML/JS files on disk, auto-discovered and registered
• MCP support: Connect to external Model Context Protocol servers and use their tools within agents
• Unified tool abstraction: Actions, agents, workflows, and commands are all ToolSpec objects — interchangeable in any context

Examples

Terminal Client

# Simple inference query
lm q "Say hello in three words"

# Generate a commit message (git plugin)
lm commit

# Execute shell command via agent (shell plugin)
lm shell "ls -la"

Node.js: Using an Agent with Tools

import { Agent, Lm } from "@agent-smith/agent";

const lm = new Lm({
    name: "openai",
    serverUrl: "http://localhost:8080/v1",
    apiKey: "",
    onToken: (t) => process.stdout.write(t),
});

const agent = new Agent({
    name: "my-agent",
    lm,
    onToken: (t) => process.stdout.write(t),
});

await agent.run("Hello, how are you?", {
    system: "You are a helpful assistant",
    tools: [weatherTool, trafficTool],
});

WebSocket Client

import { useWsServer } from "@agent-smith/wscli";

const ws = useWsServer({
    url: "ws://localhost:5184/ws",
    onToken: (t) => process.stdout.write(t),
});

await ws.connect();
await ws.executeAgent("my-agent", "Translate to German: Hello world");

Requirements

Supported inference backends:

• Llama.cpp: the first class citizen
• Any OpenAI-compatible API server (OpenAI, LM Studio, etc.)

Related Repositories

Repository	Description
agent-smith-plugins	Plugin extensions: git, sqlite, fs, shell, search, video, autodoc
agent-smith-ui	Web interface for Agent Smith

Agents processing a real-world job, thinking, working and adapting.