Language Service JavaScript Interface (LSJI)

Antigravity-level agentic autonomy meets the ultimate simplicity of self-hosting.

Overview

LSJI (Language Service JavaScript Interface) is a production-grade framework for hosting AI agents with zero infrastructure complexity. Simply provide a Gemini API key (or OpenAI/Anthropic/local) and deploy autonomous agents with:

• 🧠 LLM Agent Runtime — ReAct pattern with tools, memory, and streaming
• 🛡️ HITL (Human-in-the-Loop) — Approval gates for sends, charges, code execution
• 🔄 Durability — Checkpointing & recovery from any failure
• ♻️ Idempotency — Duplicate prevention on retries
• 💰 Budget Controls — Token/cost limits with circuit breaker
• 🌐 Web Control Panel — Real-time thought logs, approval queue, budget monitoring
• 🔌 Plugin System — Drop-in custom tools via lsji-plugins/
• 💾 Pluggable Storage — node:sqlite (built-in), better-sqlite3, or in-memory

Perfect for: Self-hosted AI assistants, autonomous coding agents, research agents, workflow automation.

Quick Start (30 seconds)

# Install
npm install @game_ryo/lsji

# Set your API key (Gemini, OpenAI, or Anthropic)
export GEMINI_API_KEY="your-gemini-key"
# or: export OPENAI_API_KEY="your-openai-key"

# Start the runtime server with web UI
npx @game_ryo/npx @game_ryo/lsji serve

# Open http://localhost:3456 — create runs, watch thought logs, approve actions

CLI Usage

# Runtime server with control panel
npx @game_ryo/lsji serve --port 3456              # Start server + UI
npx @game_ryo/lsji serve --no-ui                  # Headless mode

# Agent operations
npx @game_ryo/lsji agent run --task "Write a report on AI trends" --provider gemini --model gemini-1.5-flash
npx @game_ryo/lsji agent run-durable --task "Analyze codebase" --workflow-id my-analysis

# Budget & approvals
npx @game_ryo/lsji budget status --budgetId my-project
npx @game_ryo/lsji hitl list                       # Pending approvals
npx @game_ryo/lsji hitl approve --id <id> --reason "Approved"

# Plugins
npx @game_ryo/lsji plugin list
npx @game_ryo/lsji plugin create --name my-tools   # Generate template

# RL (legacy)
npx @game_ryo/lsji train --episodes 500
npx @game_ryo/lsji play --hand 0

Programmatic Usage

import { createLLMAgent, startServer } from 'lsji';

// Create an agent with full production features
const agent = await createLLMAgent({
  llm: { provider: 'gemini', model: 'gemini-1.5-flash', apiKey: process.env.GEMINI_API_KEY },
  budget: { maxCostPerRun: 5, maxCostPerDay: 50 },
  hitl: { enabled: true, defaultTimeout: 300000 }, // 5 min approval timeout
  memory: { conversation: true, episodic: true, semantic: true },
});

const result = await agent.run("Research TypeScript best practices and create a summary");
console.log(result.answer);

// Or start the full runtime server
const server = await startServer({ port: 3456 });
// Visit http://localhost:3456 for the web control panel

Web Control Panel

When you run npx @game_ryo/lsji serve, you get a real-time dashboard at http://localhost:3456:

| Panel | Features | |-------|----------| | Runs | List active/completed runs, start new ones | | Thought Log | Stream agent's reasoning (💭 thoughts, ⚡ actions, 👁 observations) | | Approval Queue | One-click approve/reject for file writes, API calls, emails, code exec | | Budget | Real-time cost tracking with progress bars & circuit breaker status | | Plugins | View loaded custom tools |

Tool Plugin System

Drop .js files in lsji-plugins/ to extend agent capabilities:

// lsji-plugins/my-tools.js
export default {
  name: 'my-tools',
  version: '1.0.0',
  tools: {
    my_custom_action: {
      name: 'my_custom_action',
      description: 'Do something custom',
      parameters: { type: 'object', properties: { input: { type: 'string' } }, required: ['input'] },
      requiresApproval: false,
      async execute({ input }) { return { result: `Processed: ${input}` }; }
    }
  }
};

npx @game_ryo/lsji plugin list  # Shows: my-tools (1 tools)

Architecture

src/
├── core/              # RL foundations (Env, QLearning, Agent)
├── storage/           # Pluggable backends (sqlite, better-sqlite3, memory)
├── execution/         # Production systems
│   ├── budget/        # TokenCounter, CostTracker, CircuitBreaker
│   ├── hitl/          # ApprovalGate, ApprovalStore, Notifier
│   ├── idempotency.js # Duplicate prevention
│   └── engine.js      # Durable execution with checkpoints
├── llm/               # LLM Agent Framework
│   ├── providers/     # OpenAI, Anthropic, Local (Ollama), Gemini
│   ├── llm-agent.js   # ReAct agent with tools & memory
│   ├── tools/         # Built-in: web_search, file_read, file_write, code_exec, api_call, send_email, db_query
│   ├── memory/        # Conversation, Semantic, Episodic
│   ├── plugins/       # Dynamic plugin system
│   └── prompt-manager.js
├── server/            # Runtime server + Web UI
│   ├── index.js       # Express + Socket.io
│   └── ui/            # React + Vite control panel
└── cli.js             # All commands

LLM Providers

| Provider | Models | Setup | |----------|--------|-------| | Gemini | gemini-1.5-flash, gemini-1.5-pro, gemini-2.0-flash | GEMINI_API_KEY | | OpenAI | gpt-4o, gpt-4o-mini, gpt-4-turbo | OPENAI_API_KEY | | Anthropic | claude-3-5-sonnet, claude-3-opus, claude-3-haiku | ANTHROPIC_API_KEY | | Local | Any Ollama model | OLLAMA_HOST (default: http://localhost:11434) |

Configuration

Environment variables:

• GEMINI_API_KEY / OPENAI_API_KEY / ANTHROPIC_API_KEY — API keys
• LSJI_STORAGE — Storage backend (sqlite, better-sqlite, memory, default: sqlite)
• LSJI_DB_PATH — Database file path (default: ./lsji.db)
• LSJI_SERVER_PORT — Server port (default: 3456)

Why LSJI for AI Agents?

| Feature | Typical Agent Frameworks | LSJI | |---------|-------------------------|------| | Self-hosted | ❌ Often cloud-only | ✅ Single binary, zero deps | | HITL approvals | ❌ Manual or none | ✅ Built-in with web UI | | Durability | ❌ Lose state on crash | ✅ Checkpoint & resume | | Cost control | ❌ Unlimited spend | ✅ Budget + circuit breaker | | Idempotency | ❌ Your problem | ✅ Automatic | | Web UI | ❌ Separate project | ✅ Embedded, zero config | | Plugins | ❌ Complex SDK | ✅ Drop .js files | | Gemini support | ⚠️ Often missing | ✅ First-class |

Roadmap

• [ ] v0.4 — Multi-agent orchestration, A2A protocol
• [ ] v0.5 — WASM sandbox for code_exec, vector memory
• [ ] v1.0 — Linux Foundation Sandbox, stability guarantees

License

MIT License — see LICENSE for details.

Copyright 2026 ryopc org

Built for engineers who want agentic autonomy without the infrastructure burden.
If you find LSJI useful, ⭐ the repo and consider sponsoring!