OpenClaw Local Deployment Guide: Free Access to Claude Opus 4.5 and Building an Autonomous AI Assistant

What Is OpenClaw

OpenClaw (formerly known as clowdbot and moltbot) is a personal AI assistant project that has quickly gained traction in the open-source community. Unlike typical chatbots, its core strength lies in autonomous reasoning and execution — it doesn't just answer questions, but can proactively browse the web, read pages, invoke tools, and even find workarounds when it hits a dead end.

OpenClaw falls under the "AI Agent" category, which is one of the most significant evolutionary directions in artificial intelligence today. Traditional chatbots are essentially "question-and-answer systems" — the user inputs something, the model outputs a response, and the interaction ends in one or a few turns. An AI Agent, by contrast, introduces a closed-loop cycle of "perceive → plan → execute": it can sense the environment (e.g., webpage content, tool return values), formulate multi-step plans, invoke external tools, and dynamically adjust its strategy based on results. This architecture is academically known as the ReAct (Reasoning + Acting) framework, proposed by Princeton researchers in 2022 and now the foundational design of mainstream agent systems.

This guide is based on a Windows local deployment walkthrough by a Bilibili creator. It covers the full workflow — from configuring a large language model and connecting a chat interface to testing real-world capabilities — so you can get this remarkably capable AI assistant up and running quickly.

Step 1: Configure the LLM (Recommended via Terminal)

After installing the main application, the first task is to give your assistant a "brain" — by connecting it to a large language model.

Although the official documentation supports configuration through the web interface, that interface is incomplete in practice. Running configuration commands directly in the terminal is far smoother. Once you're in the terminal, the system lists all configurable options; select models to see virtually every major LLM available.

Why Not Use the Officially Recommended Claude Opus 4.5

The official first recommendation is Claude Opus 4.5, which delivers the best results — but it comes with two obstacles for users in China:

• High cost: As a paid model, available quota is limited and pricing is steep.
• Strict network requirements: Accessing it from within China demands a reliable network environment.

Claude Opus 4.5 is Anthropic's most powerful flagship model. Anthropic was founded by former core OpenAI members and focuses on AI safety research. The Claude series is renowned for long-context processing (up to 200K tokens), high instruction-following precision, and strong safety alignment. It excels at complex reasoning, code generation, and multi-step task planning — putting it on par with GPT-4o and Gemini Ultra as one of today's top commercial models. Pricing is typically token-based, and because agent applications frequently invoke tools and generate large volumes of intermediate reasoning tokens, costs can run significantly higher than in ordinary conversation settings — which is the root cause of the "high cost" concern.

Using Google Anti-Gravity Authorization to Access Top Models for Free

The Bilibili creator shared an elegant workaround: configure Google Anti-Gravity authorization, which offers two clear advantages:

1. Free to use for any Google account holder;
2. Through this channel, you can indirectly invoke the Opus 4.5 model — essentially zero-cost access to top-tier AI capability.

Better still, if you have multiple Google accounts, you can mount all of them and rotate between them, effectively eliminating quota concerns.

Authorization and Verification Steps

Navigation path: Terminal → Select Google → Select Anti-Gravity Authorization. The system generates an authorization link — copy it in full, paste it into your browser, log in to your Google account, and confirm authorization.

Important note: After completing authorization, your browser may display "This site can't be reached" — this is completely normal. "Google Anti-Gravity" works through the OAuth 2.0 authorization flow to obtain a temporary token for accessing the Google AI API. OAuth 2.0 is the most widely used third-party authorization protocol on the internet today; users never expose their account credentials to third-party apps — they simply click "Authorize" on Google's official page to complete identity verification. The "This site can't be reached" message appears because the callback URL points to a local port (e.g., localhost). At this point, manually copy the full URL containing the authorization code back into the terminal, and the local application completes the token exchange — this is the standard "Authorization Code Flow" and is completely secure. Copy the URL from the address bar, right-click to paste it into the terminal, press Enter, and after a brief wait you should see an "Authorization successful" message.

The system will automatically set the default model to Opus 4.5 — just press Enter to confirm. Once configured, ask a test question in the web chat interface (e.g., "Who are you?") to verify the model is running correctly.

Using the dedicated status command, you can clearly view the current default model, fallback models, usage statistics, quota reset time, and a list of mounted Google accounts — all very easy to manage.

Step 2: Connect a Chat Interface for Remote Control from Your Phone

With the LLM configured, the next step is connecting a chat interface so you can issue commands from your phone at any time without being tied to your computer.

In the configuration interface, select Channels → Configure Connection. You'll find that Telegram, WhatsApp, Discord, and other major platforms are all supported. The demo uses Telegram.

Full Telegram Pairing Walkthrough

Telegram provides a robust Bot API ecosystem that allows developers to create and manage custom bots through the official "BotFather" bot. BotFather is essentially a meta-bot operated by Telegram that issues Bot Tokens — unique identity credential strings. Once a third-party application holds a token, it can receive user messages and reply in real time via either long polling or Webhook through the Telegram Bot API. OpenClaw chose Telegram as the default demo channel primarily because its Bot API onboarding process is highly standardized, cross-platform (mobile, desktop, web), and ideal as a remote command interface for an AI Agent.

Here are the pairing steps:

1. Open Telegram, search for and open a conversation with BotFather;
2. Enter the creation command and give your new bot a name — note that the name must be globally unique, so choose something distinctive;

3. Once BotFather returns the Token, copy and paste it back into the terminal to complete the connection;
4. Open the newly created bot's chat window and send any message — it will return a pairing code;
5. Paste the pairing code along with the corresponding command back into the terminal and press Enter; a success message confirms pairing is complete.

From this point on, you can communicate with your AI assistant directly through Telegram in real time.

Step 3: Testing Autonomous Agent Capabilities

After completing the full setup, the creator ran a series of capability tests — and the results were quite impressive.

Web Search and Page Reading

Asking "What's the weather like in Beijing?" returned a normal result, confirming that web access works. Asking it to read the contents of a specific webpage also worked without issue.

Autonomous Problem-Solving When Blocked

What truly demonstrates "intelligence" is how it handles obstacles. The autonomous execution behavior OpenClaw displays relies on a combined architecture of Tool Use / Function Calling and multi-step reasoning. Major LLMs (Claude, GPT-4 series, Gemini) natively support Function Calling interfaces: developers pre-define tool names, parameter formats, and descriptions, and the model can actively decide during inference "which tool to call and with what parameters." The execution result is fed back as new context to drive the next round of reasoning — enabling the Agent to progressively complete complex tasks much like a human operating a computer.

Scraping tweets from Twitter: Since Twitter requires login to view content, the assistant honestly admitted it couldn't complete the task directly but proactively offered a solution — suggesting the user log in to Twitter or provide cookies. In the web interface, you can see the full reasoning chain: it first attempts the web search tool, and upon failure (due to an unconfigured API), it pivots to an alternative approach. This demonstrates genuine autonomous reasoning and execution, not a simple pre-scripted response.

Reading a YouTube video: This task took a more winding path. The assistant first noted that a plugin needed to be installed; when the user asked it to handle that itself, it independently completed the plugin installation. It then ran into an issue where its IP was flagged as a bot — but after a brief hint from the user, it came up with its own method to bypass the detection and ultimately succeeded in reading the video content. This process perfectly illustrates the Agent's closed loop of "perceive → plan → execute → feedback": every tool call result becomes the input for the next reasoning step, with the system iterating continuously until the task is complete.

In a fun twist, the creator deliberately fed it a video that criticized OpenClaw. After successfully reading it, the assistant responded with self-deprecating humor. Throughout the process, it exhibited very human-like behavioral tendencies — including what seemed like a touch of laziness, requiring the occasional nudge from the user to keep trying.

Summary: An Open-Source AI Agent Worth Trying

This walkthrough shows that the deployment barrier for OpenClaw is not particularly high — the main challenge lies in how you connect to a large language model. The approach of using Google Anti-Gravity authorization for free access (even to Opus 4.5) offers a low-cost, high-performance solution that's accessible to users in China.

More noteworthy is its autonomous agent capability: automatically installing plugins, independently bypassing detection, offering alternative solutions when blocked — these behaviors indicate that OpenClaw has moved well beyond the realm of traditional chatbots and is evolving toward a "truly execution-oriented AI assistant." The underlying ReAct framework and Function Calling technology it relies on represent the most active frontier in AI engineering research and application today.

Of course, a healthy dose of realism is warranted: its "autonomy" still depends on appropriate user guidance, and its stability and capability boundaries need validation across more scenarios. But for anyone eager to explore personal AI assistants, OpenClaw is undoubtedly a high-quality open-source project well worth getting your hands on.

Key Takeaways