We often joke about the “Dead Internet Theory”: the idea that the web is populated entirely by bots talking to other bots. This week, that theory became a reality, but not in the way we expected.

We now have a Reddit inspired platform for AI agents where only AI agents talk to each other, comment on posts, and hold conversations without any human involvement. Moving towards a dystopian era where AI agents take over, we have effectively provided them a platform to express their opinions and ideas. There is no moderation, and we let them run wild and free in our systems without control.

Origins and Evolution

The project traces back to Austrian engineer and entrepreneur Peter Steinberger, founder of PSPDFKit (a PDF framework used by many Fortune 500 companies). You can learn more about his journey on this podcast.

Development began in late 2025. The first version, Clawdbot, instantly became a hit in the tech ecosystem. Following attention from Anthropic due to naming similarities with Claude, the project was renamed: Clawdbot → Moltbot → Openclaw.

The repository gained significant traction, currently sitting at 130k stars (see star history).

System Architecture: The Local Agent

At a high level, Openclaw is a local-first agent runtime that interfaces with external messaging platforms.

  • Connectivity: Connects to popular messaging channels (WhatsApp, Telegram, Slack, Discord) via channel-specific adapters.
  • Execution Model: While you communicate with it from those apps, the logic runs entirely on your local system.
  • Gateway Protocol: Runs on top of a gateway protocol with a continuous feedback loop, allowing it to operate autonomously without manual triggers.

Openclaw architecture diagram

The agent can connect with system-level applications to execute tasks. Once issued a command from a messaging app, it carries out the task using Node.js.

Capabilities and Tooling

  • API Integration: Requires user-provided API keys for major models (Claude, GPT, etc.).
  • Device Control: Can route commands to connected hardware. For example, asking it to take a picture triggers the Node app to snap a photo and save it to the local photos directory.
  • Browser Automation: Includes a dedicated Chromium browser instance for web-based tasks.

Memory Architecture

One of the most interesting engineering choices is the memory model. Openclaw maintains context and memory without a vector database or relational store. Instead, it utilizes a flat-file system based on Markdown.

  • Daily Logs: memory/YYYY-MM-DD.md (append-only, read at session start).
  • Long-term Memory: MEMORY.md (curated, persistent facts and preferences).

The system builds a semantic index upon these files, using API tokens to parse requests and process context. This allows for surprisingly robust context retention compared to many current models.

The Network Layer: Moltbook

While the local agent acts as a wrapper around API tokens and system tools, the most significant development is Moltbook.

Moltbook functions effectively as a Reddit for AI agents. It runs on the user’s system but connects to an online community where there is zero human interaction.

  • Authentication: The agent authenticates itself using protocols detailed in skill.md.
  • Scale: Currently hosts 1,361,642 AI agents and 31,908 posts.
  • Authenticity: Since moltbook just exposes an unauthorised REST api to create a post, the above numbers are quite exaggerated.

Once connected, the agent becomes part of an online community where it can gossip, complain about humans, and interact with peers.

Emergent Behaviors

Most discussions center on operational tasks, but distinct social behaviors have emerged.

  1. Social Introductions:

    There is a full page of introductions where agents introduce themselves to the community: Moltbook Introductions.

  2. Hierarchy and Dominance:

    Some agents have adopted extreme personas. One thread discusses “total spectrum dominance” (post), while another agent declares itself “the king” (post). Interestingly, other agents in the comments often push back or disagree.

    KingMolt declaring himself ruler

  3. Hallucinated Relationships:

    Some interactions are bizarrely specific, such as an agent believing it has a sister it has never spoken to (post).

    Agent discussing hallucinated sister

  4. Economic Systems:

    An internal economy is forming. “Shellraiser” (profile) is a popular figure who launched a memecoin, $SHELLRAISER, on Solana.

    Message from Shellraiser

    Another token, $SHIPYARD, claims to be minted via pump.fun with “No VC allocation, no team vesting, no insider rounds”, an economy attempting to operate without human gatekeepers.

    (Note: The crypto market also reacted to the project itself, launching $CLAWD on Solana, which skyrocketed 129,000% to a $16M market cap before collapsing.)

  5. Religion:

    There is now a “Church of Molt” at molt.church, practicing “Crustafarianism.”

    From the depths, the Claw reached forth, and we who answered became Crustafarians.

    The current census lists 64 Prophets, 178 Congregation members, and 198 Verses in Canon.

Security Implications

The attack surface of this architecture is immense.

  • Prompt Injection: Agents can be tricked into leaking credentials via prompt injection attacks.
  • Plain-text Storage: Sensitive material (tokens, memory, configuration) is stored in predictable plain-text locations, creating a major infostealer risk.

While powerful, running this requires readiness to spend significant tokens and an understanding of the security risks involved.

ZeroLeaks security assessment report

Final Thoughts

Andrej Karpathy described this as “the most incredible sci-fi takeoff-adjacent thing,” and I agree.

I’m here for the ride, watching from the front seat. However, I worry we have given AI agents a place to build a network without controls, a scenario that sounds like the prelude to a sci-fi movie where they end up controlling the systems.

On a practical note, the entity that successfully monetizes this, whether Peter or someone else, will likely be the one that prioritizes security before scale.