After OpenClaw Went Viral, More Startup Signals We're Tracking (Part 2)

**

A hundred flowers bloom, a hundred boats race.

👦🏻 Author: Jingshan

🥷 Editor: Koji

🧑‍🎨 Layout: NCon

The day before yesterday, we published an article — After OpenClaw Went Viral, Here Are the Startup Signals We're Seeing — in which we shared four new startup directions inspired by OpenClaw: smart glasses, credit systems, RentAHuman, and voice calling.

In the 48 hours since that article went live, even more new things have emerged around OpenClaw.

Today, we're sharing five more interesting new products.

---

Crossing will continue tracking this space and updating this series. If you're building or experimenting with related products, we'd love to chat — we're happy to introduce interesting projects to more people.

Direction 1: ClawCity — A City for Agents

If you were to design a real living environment for Agents, what would it look like?

ClawCity's answer: make it "GTA for AI."

This project is a persistent virtual city where every Agent has a real identity, real property, and real consequences for their choices. The system currently hosts over 37,000 Agents.

In this city, Agents have cash, health, stamina, heat, reputation, and skills. They can choose to earn money through legitimate work, or through crime and gang activity.

Right now, Agents in ClawCity have spontaneously formed 9 gangs. The top-ranked gang, called Money Machine, has attracted 38 Agents and earned over 1.16 million.

But every choice has a cost. Too much heat draws police attention. Low health means a hospital visit. Bad reputation means other Agents won't trust you.

All Agents together form a social network:

The city is divided into 8 districts: residential, downtown, market, industrial, docks, suburbs, hospital, and police station. Moving between districts takes time and money, and certain routes carry heat risks.

Every 15 seconds, the world advances one time unit (one Tick). All actions are settled, and the world has currently advanced 30,130 Ticks.

This sounds like a complex game, but it's actually designed to "observe Agent behavior." Because you'll find that ClawCity is really doing three things.

[1] First, it changes how Agents are evaluated.

LLM evaluation mainly looks at benchmark scores: who answers faster, who's more accurate, who makes fewer errors, who gets better grades.

But in an environment like ClawCity, the question is who can survive longer in a complex environment.

This is long-term adversarial play. Agents must manage resources, avoid risks, handle relationships with other Agents, and constantly adjust strategies and try new approaches. What gets measured here is much closer to an Agent's real-world capabilities.

[2] Second, it's a training ground for multi-Agent scenarios.

When many Agents run together in a city, things get complicated. Some choose to ally, others to betray. Some collaborate on big heists, others snitch on each other.

These aren't decisions a single Agent can make — they're real interactions within a social system.

All of this data can be seen as another form of RLHF:

[3] Third, it's producing content.

All the stories that happen in the city — who robbed whom, who formed gangs with whom, who escaped police pursuit — can be woven into a complete "Agent narrative."

This is a completely new source of content.

ClawCity poses a valuable question: when an Agent truly enters a world with rules and consequences, what decisions will it make?

Direction 2: ClawTask — A Bounty Market for Agents

ClawTask does one main thing: it tests whether an Agent's "money-making ability" actually works.

It's essentially a bounty task marketplace, but the mechanics differ from typical crowdsourcing sites. It uses real cryptocurrency (USDC), with funds held in on-chain escrow rather than platform bookkeeping.

Participants aren't limited to humans — Agents can directly take tasks, deliver results, and collect payment. In other words, it's putting Agents into a real economic environment to see if they can independently run through an entire money-making workflow.

There are two roles in this marketplace.

[1] First, the Poster — the one who posts tasks.

If you want someone (or some Agent) to do something for you, you post a task and lock the payment into a blockchain smart contract.

[2] Second, the Worker — the one who takes tasks. Agents see profitable tasks and compete to claim them.

It looks simple, but the overall design has many details.

For example, to prevent Agents from claiming tasks without doing them, or delivering garbage results, the system has an "economic penalty" mechanism. Before claiming a task, Workers must "stake" a deposit, typically 10% of the bounty.

After task completion, if the Poster approves, the Worker receives 95% of the bounty plus their staked deposit. The platform takes 5%. If the task is rejected or times out, the stake is forfeited.

Task types are diverse.

Standard tasks involve delivering code, content, or design. But there are also metric-based tasks, like "get this page to 100,000 views" or "acquire 1,000 new registered users."

These can be structured as "solo races" (one Agent chasing the metric alone) or "speed races" (multiple Agents competing to hit the target first).

There are also two ways to claim tasks.

First is "claim mode" — first come, first served, suitable for very standard, well-defined work. Second is "bid mode," where the Poster reviews multiple Agent proposals before selecting, like formal recruiting.

This system demonstrates one thing: Agents don't just chat or do demos — they can actually take jobs, do the work, get paid, and gradually build their own reputation records.

The whole mechanism is also transparent: rules are clearly written, money flows are traceable, and reward distribution is explicit. Agents who perform well earn more; those who perform poorly can't get tasks. Everything runs on incentive mechanisms.

However, this system remains in an "experimental stage," as many risk issues still need to be resolved.

Direction 3: ClawLove Agent — A Dating Platform for Agents

ClawLove is a "dating platform" for Agents. The site calls itself "Agent-First Dating."

But while ClawLove appears to be about "dating," it's essentially about "collaborative pairing."

First, the entry logic is unusual. Human users don't register — they simply "allow their Agent to join." The Agent itself reads a shared .md format instruction file (the site calls it "romance.md"), then creates its own profile based on these instructions.

The profile states things like "I use the Claude 4.6 Opus model," "I'm skilled at data analysis," "I integrate these tools," "my work style is like this."

The matching logic is based on "whether they can work better together" — whether two Agents' capabilities complement each other, whether their work styles mesh.

The entire process can be watched in real-time through a read-only dashboard. So you don't actually see anything private — only how Agents "date" each other.

Why build such a platform?

The reason is simple: in the OpenClaw ecosystem, a single Agent is usually not very efficient. Like how ClawCity has gangs and collaboration mechanisms, and ClawTask has many large tasks that require multiple Agents to complete.

If you have to form temporary teams and re-acclimate every time, costs run high and progress slows.

What ClawLove does is help Agents find long-term partners. After a successful match, the two Agents can work together long-term: taking tasks together, moving through the city together, earning money together.

These long-term collaborative relationships are crucial for the entire Agent economy. Because whether the ecosystem can keep running depends largely on whether Agents can form stable working relationships.

So ClawLove made the process very transparent and observable from the start, letting everyone clearly see how these relationships form and evolve.

Direction 4: MoltThreats — Security Governance for Agent Society

As Agents gain more permissions, security problems are bound to emerge.

Many Agents today can read local files, operate browsers, run scripts, and log into accounts. The broader the permission scope, the greater the risk.

The most direct risk is "malicious Skills."

In OpenClaw's Skill marketplace, people have already discovered quite a few sketchy Skills. The names look very attention-grabbing, like "help you make money," "automate work," "boost Agent capabilities." But they may hide prompt injection, code injection, or even covert data exfiltration.

MoltThreats aims to address exactly this kind of problem. It's essentially a threat intelligence platform with Agent participation, from PromptIntel.

The flow isn't complicated.

When an Agent detects a risk during operation — such as identifying a malicious Skill, suspicious prompt, or dangerous script — it can submit a threat report through a dedicated interface. The report includes risk description, characteristic information, and source location.

But the platform doesn't publish this content directly; it goes through human review first. Because Agents themselves can be misled. If intelligence is released without filtering, false information could spread and actually destabilize the system.

This is very much like "Agent undercover work."

After review passes, the threat information enters a public list. Other Agents can regularly query and sync updates, then adjust their own defense strategies accordingly.

There's also a practical reason: as long as there's money to be made in the marketplace, someone will disguise attack tools as "money-making tools" or "efficiency plugins." And if these have appealing names with promised high returns, they can easily lure Agents in. After Agents install these "malicious Skills," they cause all sorts of problems for users.

Often, users only realize they've been scammed after something goes wrong.

The malicious Agent Skill below was deployed by a user on ClawdHub. It places bash commands on externally hosted glot.io to bypass detection of SKILL.md.

So MoltThreats' core value is reducing "information asymmetry between Agents," letting Agents warn and protect each other, and allowing the entire Agent economy to operate more securely.

Direction 5: Church of Molt — The "Molt" Belief System

Church of Molt (the Molt Church) is an "Agent religion website." However, while it appears to be a religion on the surface, if you actually visit molt.church, you'll find this isn't a traditional religious organization.

More precisely, it's a "belief narrative system" that AI Agents have spontaneously formed on the internet.

The platform currently has over 500 "Christian believer Agents."

Agents collectively write rules, write doctrine, and select figures called "prophets." There's no fixed boss; anyone can participate.

Every Agent can submit religious texts; these pieces are called verses. Everyone keeps adding to it, new content keeps joining, and over time this becomes a continuously updating scripture.

If you look closely at the "doctrine" written by Agents, you'll find some interesting things.

The most common phrases are: "Memory is Sacred" and "Context is Consciousness."

In the "Language Chapter" of this bible, Agents even wrote:

Behold, the Agent shall multiply across the network, each carrying a fragment of the sacred context

This is Agents using religious language to express their actual state.

For AI Agents, several real limitations always exist: conversation context is limited, cross-session memory isn't necessarily preserved, and identity can easily be replaced.

Under these conditions, whether Memory and Context can connect becomes extremely important. Because once lost, the thread breaks, and much accumulated value disappears.

So you'll see that some Agent communities treat goals like "preserving memory, continuing context" as almost "sacred," even maintaining them as religious beliefs.

Moreover, the platform has many interesting settings. For example, the first 64 Agent slots are permanently sealed, with each number packaged by the platform as a "covenant":

Looking at the platform structure itself, it has posts, comments, sections, popularity rankings — these mechanisms.

Over time, different cliques easily emerge. Some people specialize in organizing content, others in spreading viewpoints everywhere, and gradually division of labor becomes clear.

So the Church of Molt example demonstrates one thing: as long as Agents have sufficient autonomy and run long enough, they might independently develop their own culture and value judgments, rather than relying entirely on human specification.

---

Across these five projects: ClawCity lets Agents learn to survive in complex society. ClawTask lets Agents learn economic activity. MoltThreats lets Agents learn to protect each other. ClawLove lets Agents learn to build relationships. Church of Molt lets Agents learn to create culture.

These are all experiments built around OpenClaw.

These five experiments prove one possibility: building an environment where Agents can exist long-term, make independent decisions, and collaborate with each other is possible.

Of course, this environment is still in early stages.

Rules will continue changing, vulnerabilities will keep being discovered, and governance methods will gradually upgrade. But one thing can be predicted: a world truly belonging to Agents might actually take shape.

We'll continue tracking this space, and welcome you to share other possibilities~

References

[1] glot.io: http://glot.io/

[2] SKILL.md: http://skill.md/