OpenClaw Starter Kit: GPT 5.2 Multi-Agent Framework

Purpose and Significance

OpenClaw Starter Kit provides a production-ready foundation for building autonomous AI agents with GPT 5.2, designed for developers and researchers who need specialized, task-focused agents without lengthy setup complexity. By bundling essential components—a main orchestration agent, sandboxed dev and research workers, persistent knowledge graphs, and Telegram integration—this kit accelerates deployment of multi-agent systems for real-world applications like code automation, research synthesis, and knowledge management.

Key Features

• GPT 5.2 Integration — Leverages the latest OpenAI model as the default LLM backbone for reasoning and coordination.
• Multi-Agent Architecture — Main agent orchestrates specialized workers: Dev Agent for sandboxed coding tasks, Research Agent for web browsing and information retrieval.
• Sandbox Isolation — Dev agent runs in Docker with filesystem restrictions (workspace-only) and disabled network access; Research agent operates with network access but is isolated from private files.
• Internal Knowledge Graph — Markdown-based persistent memory system (workspace/kg/) for maintaining context across sessions, with INDEX.md as entry point and inter-linked topic notes.
• Telegram Chat Interface — Direct messaging bot for real-time interaction without requiring web UI setup.
• Configuration as Code — JSON-based agent definitions and tool policies; environment variables for sensitive credentials (OPENAI_API_KEY, bot tokens).
• Workspace Protocol — Built-in ORCHESTRATION.md defines patterns for serious project work and agent coordination.
• One-Command Setup — ./setup.sh bootstraps the entire stack with prerequisite validation (Node.js 20+, Docker, OpenAI API key).

Getting Started

Prerequisites: Node.js 20+, OpenAI API key, Telegram bot token (from @BotFather), and optionally Docker for sandboxed execution.

Quick Setup (5 minutes):

1. Install OpenClaw CLI: npm install -g @anthropic/openclaw
2. Initialize config: openclaw init
3. Set environment variables (OPENAI_API_KEY, TELEGRAM_BOT_TOKEN, TELEGRAM_ALLOW_FROM)
4. Run bootstrap: ./setup.sh
5. Start gateway: openclaw gateway start
6. Message your Telegram bot—it responds immediately.

For manual setup without ./setup.sh, copy openclaw.example.json to ~/.openclaw/openclaw.json, create ~/.openclaw/.env with your API key, and customize the config keys under channels.telegram and agents.list.

Core Architecture

Main Agent: Your primary assistant that handles direct requests and spawns workers for specialized tasks.

Dev Agent: Executes coding tasks in a sandbox with file access limited to the workspace directory and shell execution available. Useful for code generation, testing, and automation. Trigger with "spawn a dev agent to [task]" or automatic invocation for complex coding requests.

Research Agent: Performs web searches, page fetching, and browser automation. Network-enabled but logically isolated from private file systems. Invoked via "research [query]" commands; output is treated as "tainted" (external data).

Knowledge Graph: Workspace-based markdown system where INDEX.md serves as the entry point. Agents automatically read and update topic notes (kg/topic-name.md) to maintain persistent context, perfect for building institutional memory during long-running projects.

Configuration and Customization

The starter kit targets OpenClaw v2026.2.x. Most configuration happens via environment variables or the generated ~/.openclaw/openclaw.json. Key customization points:

• Edit SOUL.md in your workspace to define agent personality and tone.
• Modify agent tool policies in the config's agents.list to allow/deny specific tools per worker.
• Expand the knowledge graph by adding topic notes and cross-linking them for semantic navigation.
• Enable or disable sandboxing via workspace, sandbox, and tool_policy fields per agent.
• Store all secrets in ~/.openclaw/.env (chmod 600) rather than in JSON files.

Troubleshooting

• Bot doesn't respond: Verify openclaw gateway status, check TELEGRAM_BOT_TOKEN is correct, and ensure TELEGRAM_ALLOW_FROM contains your numeric user ID (get it from @userinfobot).
• Code execution fails: Install Docker (docker --version), confirm the daemon is running, and re-run ./setup.sh to enable sandboxing.
• Research agent can't browse: Verify network access is enabled in the research agent config and that web_search and web_fetch are in the allowed tools list.

Who It's For

AI engineers and research teams building multi-agent systems for code automation, document analysis, or knowledge management. DevOps and platform teams automating infrastructure tasks with safe sandboxed execution. Independent researchers prototyping autonomous workflows with persistent memory. Startups bootstrapping AI-native products without building agent orchestration from scratch.

Next Steps

• Customize SOUL.md to give your agent a unique personality and working style.
• Build out your knowledge graph by adding topic notes as you work with the agent.
• Install recommended skills (see skills.md) to extend agent capabilities.
• Explore the ORCHESTRATION.md protocol for managing complex, multi-session projects.

Resources

• GitHub Repository — Source code, example config, and quick-start scripts.
• OpenClaw Documentation — Full schema reference and best practices OpenClaw.
• OpenClaw Discord Community — Real-time support and agent architecture discussions.

Source: GitHub repository casparfelten/openclaw-starter, updated February 2026.