Another AI app goes viral! The prototype of a personal AI assistant: Clawdbot is here.
An AI assistant capable of self-improvement that runs on a local computer is changing how people perceive digital assistants. The open-source project Clawdbot demonstrates new possibilities for personal AI assistants: not only can it remember user preferences, control smart home devices, and manage schedules, but it can also independently learn new skills by accessing the computer’s file system and terminal commands, and can even rewrite its own code to meet user needs.
Federico Viticci, author of the well-known tech blog Macstories, recently wrote that this project, developed by Peter Steinberger, has quickly gained popularity in the AI community. Unlike mainstream applications such as ChatGPT or Claude, Clawdbot runs entirely on the user's local device and interacts via common messaging software like Telegram and iMessage, without the need to install additional apps. Users can converse with it using voice or text, and it intelligently selects the reply format based on the interaction method.
The article points out that Clawdbot's core breakthrough lies in its “malleability.” It stores settings, preferences, and memories in Markdown documents within a local folder, allowing users to view and modify them directly, or have it autonomously adjust functions through conversation. This design transforms AI assistants from closed products into open platforms, enabling unlimited customization to suit personal needs.
This project has sparked discussion about the future of AI assistant form factors. As AI becomes capable of creating any function on demand, the roles of traditional app stores and independent developers are set to be redefined. Although Clawdbot remains a niche technical project, the trend it represents—highly personalized and adaptive AI software—may signal a profound transformation in the software industry.
Locally Run Intelligent Agent
Clawdbot's architectural design shows a fundamental difference from mainstream AI applications. It consists of two main components: an LLM-driven agent running on the user's computer, and a "gateway" system that connects to multiple messaging apps.
The project supports a variety of mainstream models, including Claude and Gemini, but all settings, user memories, and commands are stored locally as folders and Markdown documents.
This design is similar to the philosophy of note-taking app Obsidian: while it relies on cloud-based LLM services for intelligent capabilities, data storage and operational logic remain entirely local, giving users complete control.
In the article, Federico Viticci states, most critically, Clawdbot is granted access to the computer's shell and file system. Once authorized, it can execute terminal commands, instantly write and run scripts, install new skill modules, and configure the MCP server for external integration capabilities. This architecture transforms the AI assistant from a passive response tool into a proactive task execution agent.
The article notes that Federico Viticci used 180 million API tokens during testing. He named his Clawdbot assistant "Navi" and interacted with it through Telegram. Navi, running on his M4 Mac mini server, is able to control Spotify, Sonos speakers, Philips Hue lights, and his Gmail account.
Self-Evolving Capability
Clawdbot's most striking feature is its ability to self-improve. Users can directly request it to add new functions, and it independently completes the entire development process.
In one test, Federico Viticci asked Clawdbot to add image generation capabilities. The assistant quickly configured Google's Nano Banana Pro model support and even proactively explained how to securely store Gemini credentials in macOS Keychain. Afterwards, Federico Viticci asked it to create an avatar combining the original mascot and Navi from The Legend of Zelda. Clawdbot found relevant materials via Google search and generated the image.
This capability extends to more complex scenarios. Previously, Viticci had developed a Shortcut for Club MacStories members that used a Groq-hosted Whisper model to transcribe audio. He sent the related article link to Clawdbot, asking it to add a skill for transcribing Telegram voice messages. Two minutes later, Clawdbot created a new skill module adapted from the original Shortcut.
To enable more natural interaction, Viticci further requested that the assistant intelligently match the reply format: voice requests receive voice replies, text requests receive text replies. Clawdbot autonomously researched ElevenLabs' latest TTS model documentation, requested API credentials, and created three test voices with different personalities for selection. Ultimately, Navi gained voice reply capabilities, supporting Italian, English, or a mix of both languages.
Replacing Traditional Automation Services
Clawdbot demonstrates the possibility of replacing cloud-based automation services with local AI agents. Viticci experimented with using it to substitute some Zapier automation flows that have been running for years to save subscription costs.
One example is automatically creating projects for MacStories Weekly. The original Zapier set-up would, after sending the newsletter every Friday, check the RSS feed, increment the issue number, and use the Todoist API to create a new project.
The article notes that Federico Viticci asked Clawdbot if it could replicate this function, and it proposed a solution: set up a cron job on the Mac mini to check the RSS feed every few hours, automatically creating a project when a new issue is found.
After five minutes of conversation, Clawdbot completed all the setup on the Mac, without relying on cloud services or requiring a subscription. This made Federico Viticci think about how many layers of automation and services could be replaced by simply giving Clawdbot prompts and shell access.
Every morning, Clawdbot automatically sends daily reports based on data from Calendar, Notion, and Todoist, along with illustrations generated daily by Nano Banana. All of these functions are implemented via local cron jobs, completely independent from third-party automation platforms.
Clawdbot's memory system is also based on local files. It automatically generates Markdown logs each day to record interactions; these files can be imported directly into Obsidian, searched with Raycast, or automatically processed with Hazel.
Challenging the Software Ecosystem
Clawdbot sparks deeper discussions about the future of app development. As AI assistants become able to create any function on demand, the value proposition of traditional app stores and independent developers will be challenged.
The article states that during testing, Federico Viticci had Clawdbot create a virtual remote for his LG TV and set up a personalized voice morning briefing—all accomplished with simple text commands. He poses the question:
If an AI assistant can precisely create features to suit individual needs, why go to the App Store for prefabricated solutions developed by others? When any automation need can be fulfilled by sending a message to a digital assistant, what is the role of tools like Shortcuts?
OpenAI's app business CEO Fidji Simo once wrote that AI labs should make more use of model capabilities to build personal super assistants, addressing the “overcapacity” issue. Clawdbot is a practical embodiment of this idea: it is not limited by pre-set developer functions, users can determine the boundaries of the assistant’s capabilities, and check the background logic at any time.
Viticci states that using Clawdbot taught him more about SSH, cron jobs, web APIs, and Tailscale in one week than in nearly two decades of computer experience. This experience of getting the computer to execute any task through conversation is both exciting and thought-provoking.
The article points out that the project also shows the potential of contemporary AI agents once granted appropriate permissions: they can build tools and become smarter for specific users via quasi-recursive improvement. All major AI companies have noticed this trend, and recent feature releases have focused on virtual file system sandboxes and command line access.
Risk Warning and DisclaimerThe market has risks, and investment should be made cautiously. This article does not constitute personal investment advice, nor does it take into account the individual investment goals, financial situation, or needs of any particular user. Users should consider whether any opinions, views, or conclusions contained in this article are suitable for their own circumstances. Any investment made on this basis is at one's own risk.
