Background and the Interesting Origins of OpenClaw
I want to share some informal insights regarding my three weeks of observing and using OpenClaw. The project’s naming is quite interesting; it was originally called moltbot but was changed to OpenClaw due to a name conflict.
There is a sci-fi background here: the first chapter of the novel Accelerando is titled “Lobster,” which describes humans uploading lobster consciousness to the cloud, where they gain self-awareness and become virtual creators. While I am unsure if the founder intended this as a tribute, modern AI scenarios certainly make such sci-fi plots feel much closer to reality.
I named my own robot Melonclaw. My setting for it is a robot with a sharp “Claw” exterior but a heart as sweet as a “Melon”—a very personality-driven assistant.
Core Architecture: Long-term Memory and “Heartbeat”
The innovation of OpenClaw is mainly reflected in two aspects:
- Long-term Memory: Its implementation logic is actually quite simple—it writes everything into Markdown files. However, this gives it a fundamental difference from ordinary Agents. While a standard Agent starts as a blank slate with every new session, OpenClaw possesses continuous memory.
- Heartbeat Mechanism: By default, it runs tasks (Cron Jobs) every 30 minutes. This transforms it from a “passive response” system into one capable of “active interaction”. The founder once shared an example: while he was hospitalized, the Agent proactively sent messages of concern because it monitored his calendar or email information. This creates a very different user experience.
My Hardware and Security Configuration
For hardware, I used one of my old 2018 MacBooks. Regarding the setup, I chose Discord as the primary communication portal.
- Platform Choice: Compared to WhatsApp or Telegram, which require phone numbers, Discord allows for easy account creation and has a very stable Bot API. I highly recommend starting here.
- Security: To ensure safety, I did not run the program on my primary computer; instead, I used a “retired” old laptop as a server. I created a specific user account without system administrator privileges to run the program and restricted its access to a specific folder to prevent it from reading my private data.
- Pairing Mechanism: The first message sent to the robot requires a Pairing Code that must be manually approved in the backend, ensuring that others cannot call your robot at will.
Practical Cases: Automated News and Code Development
I initially had it summarize AI news, community updates, and macroeconomic trends daily. However, it is important to note that it does not have a built-in search function by default. I later configured a Brave Search API Key, which significantly improved its search quality.
A more challenging attempt involved writing code. I had a requirement to export Discord chat records into a specific text format.
- I provided it with a standalone GitHub account and an open-source framework.
- It wrote the code, called code review tools, fixed bugs, and even recorded a video demonstrating the functionality to confirm the feature was complete.
- It even learned to use Playwright to simulate browser operations and loaded plugins to bypass Discord’s bot verification. This potential for self-evolution is very exciting.
Model Selection and High Costs
One of my biggest realizations over these three weeks is: it is truly useful, but also truly expensive.
- Cost Consumption: I used Gemini Pro. Initially, I had a $300 credit from GCP, but it was burned through in just two weeks, so now I am paying out of pocket. Because the “heartbeat mechanism” reads files and executes tasks every 30 minutes, it is extremely token-intensive.
- Model Trade-offs: I tried using the cheaper Gemini Flash for heartbeat tasks, but found it could not handle complex logic and tended to hallucinate. Currently, the best performer on the “Lobster Benchmark” is Claude Opus, but it is even more expensive than Google’s models. You must determine if your project’s value justifies these API expenditures.
Reflections on Future Organizational Forms
Observing the OpenClaw project, you realize the founder, Peter, pushed it forward almost single-handedly; at his peak, he made over a thousand commits in a single day. This leaves a lot to the imagination: a small team that understands AI could have productivity 10 or even 100 times that of a traditional team.
I now view the Agent as a Remote Employee. It is highly capable, but the key to its competence lies in your ability as a manager to provide enough context and useful tools. Of course, this model brings new risks: when you grant it more permissions (such as email, wallets, or API keys), the severity of any potential accident increases. How to establish this new type of trust and alignment with Agents will be a very interesting topic in the coming years.
