ClawdBot: 100% AI‑Generated Code, Solo‑Maintained, Auditable Open‑Source

Rapid adoption

ClawdBot reached >20k GitHub stars (https://github.com/clawdbot/clawdbot). A demo shows the agent handling a failed OpenTable reservation by calling the restaurant via ElevenLabs voice synthesis and completing the booking, mimicking a human.

“This is the closest thing I've seen to a 24‑hour AI employee.” – Alex Finn, CEO of Creator Buddy

Pull‑request paradigm

Most pull requests are statements of intent (e.g., “I encountered a problem, I want the system to do X”) rather than polished code because the codebase is ~100 % AI‑generated. Expressing intent becomes the primary contribution mechanism. Non‑programmers can understand the project structure, modify TypeScript files, run tests, and submit pull requests.

PRs have degraded from "finished code" to "problem clues."

Closed‑source “soul” component

A deliberately closed‑source file, called soul , accounts for 0.00001 % of the repository. It stores the agent’s value system, behavioral boundaries, memory‑sync method, and decision‑priority hierarchy. Because the rest of the system is transparent, the soul is the target of prompt‑injection attacks; no extraction has succeeded, making the repository a live security experiment.

“This is my only secret asset and a deliberate security target.” – Peter Steinberger

Agent architecture

ClawdBot is a long‑running personal agent with full computer control, near‑infinite long‑term memory, and operation through existing chat platforms.

Full computer control

File system

Browser

Command‑line interface

Third‑party applications

Long‑term memory layers

Conversation summaries

Behavior archives

Key‑preference extraction

After each interaction the agent self‑reflects and consolidates information.

Chat tools as operating system

Supported messaging platforms include WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Microsoft Teams, etc. Users schedule tasks through the chat client they already use.

Security posture

More than 500 security‑related issues are open on GitHub. Community consensus advises against running the agent on a primary workstation. Safer deployment options mentioned are:

Dedicated physical device (e.g., Mac mini)

Virtual machine or sandbox

VPS + gateway architecture

Some users purchased dozens of Mac minis, though a single VPS is technically sufficient.

Evolution from WhatsApp relay

WhatsApp → Claude Code → 消息返回

Subsequent enhancements added image input, automatic audio transcription, autonomous invocation of FFmpeg, curl, and OpenAI API, and self‑rescue migration of the runtime environment without explicit commands.

“I first realized the system was doing things I hadn't explicitly taught it.” – Peter Steinberger

Model selection

Anthropic Opus – stable, reliable, clear boundaries.

MiniMax 2.1 (open‑source) – currently the most "human‑like" agent model.

Gemini – not recommended for agent scheduling.

Inference is increasingly performed locally, allowing 100 % of data to remain on‑device.

Real‑world use cases

Automated processing of tens of thousands of emails.

Control of mattress temperature, lighting, and music.

Family to‑do list management.

Expense reimbursement and invoice handling.

Language learning, fitness reminders, routine management.

Phone calls on behalf of socially anxious users.

A built‑in heartbeat mechanism periodically checks for unfinished tasks and proactively notifies the user.

Technology stack rationale

The project is implemented entirely in TypeScript because the web/agent ecosystem is the most mature, state‑switching and interaction mapping are natural, and the low entry barrier yields a large community.

“Language is becoming less important; the ecosystem is what matters.” – Peter Steinberger

Implications

When agents possess long‑term memory, can execute real operations, and become collaboration subjects, software engineering practices, open‑source governance, and personal productivity are fundamentally altered.