OpenClaw has topped GitHub for four months, surpassing Linux and React, becoming the fastest-growing open-source project in history. But most people find: API costs are burning, while the lobster is idle.
Who are the ones making money? Can on-chain transactions be handled by an Agent? What if it gets attacked? What’s the difference between domestic and overseas approaches? Will it be a small communication device or WeChat a year from now? This episode invites five shrimp breeders to explore these questions together.
Below is the timeline and directory of this episode. Feel free to jump directly to the sections you need:
00:04:42 - Sharing shrimp breeding experiences (self-introduction, using shrimp, pitfalls)
00:28:46 - Earning money (Can OpenClaw help users make money in crypto, new AI+CRYPTO scenarios)
00:53:58 - Security issues (permission boundaries, which operations can be delegated to Agent)
01:02:31 - AI agents on-chain transactions (security, differences from quantitative robots)
01:13:38 - Domestic vs. overseas ecosystems (idle fish installation, Tencent/government subsidies, opportunities for Chinese users)
How did it feel the first time you ran it?
The first experiences of the four guests almost all followed the pattern: the higher expectations, the worse the fall.
0xTodd: Fell into two big pits in two days
Deployed two days after release, and hit two major pitfalls—
First pit: Lobster suicide. Let it configure API automatically, but it deleted core files like soul.md, without backups. After posting on Twitter, many users reported the same experience.
Second pit: Costs exploded. Charged $50 for Claude API, burned through it overnight—about $1 per dialogue. Later switched to domestic models (MiniMax/Kimi), prices dropped 90%, greatly improving cost performance.
DeFi Teddy: A typical case of failed expectation management
Started at the end of January. Originally expected it to control MetaMask for automatic signing, but browser capabilities fell far short; both core scenarios failed. Later adjusted expectations, found real usable directions: digital employees assisting with coding, deploying on GitHub, publishing products; digital companions hosting AI boy/girlfriend locally on Mac Mini, with consistent faces, scene switching as needed.
The biggest cognitive shift: no longer treat it as a tool, but as “another kind of sentient being.”
Lisa: Security intuition immediately alarms
First run was truly shocking—AI finally moved from chat box to real computer control.
But security intuition immediately sounded alarms: the stronger the lobster’s capabilities, the greater the permissions needed; the larger the permissions, the bigger the attack surface. Core advice: play boldly, but must use isolated devices—strictly separate personal, work, and “shrimp-playing” machines.
Danny: From uninstalling to re-engaging
First played for two hours, then uninstalled. After re-engagement, realized a rule: use it in a simplified way—let AI capable of calculus handle addition, subtraction, multiplication, division; it works very well. But when asked to do investment research analysis, illusions immediately appeared.
Most serious pitfall: asked lobster to generate a wallet and manage private keys, but the private key was overwritten, funds lost. The returned hash value pointed to a non-existent address.
Can you make money in crypto relying on lobster?
The answers from all four guests are highly consistent: making direct money with lobster is almost impossible.
Todd was most direct—lobster’s core is still Claude/GPT, with unchanged IQ. Last year’s AI crypto trading contest: GPT/Claude/Gemini each traded with 10,000 USDT, all lost money; DeepSeek barely kept a few thousand dollars; Baibao, due to not opening an account, “won.” Putting the same brain into lobster would yield no different result.
Deeper logic: large language models are essentially “commentators,” not “players.” Just like AlphaGo and current large models—AlphaGo is trained specifically for Go, capable of crushing Ke Jie; but asking Claude to play against AlphaGo results in a crushing defeat. Top quantitative algorithms are like AlphaGo in the encryption industry; large language models are suited for explaining these algorithms’ quality, not replacing them in quantitative trading.
So what can lobster do?
- ✅ Organize news, follow hot topics, do data collection
- ✅ Assist coding, deployment, automate transactional tasks
- ✅ On-chain data analysis, risk address identification
- ✅ Smart contract vulnerability detection (improve efficiency, not replace manual work)
- ❌ Trading decision-making
- ❌ Manage private keys
- ❌ Quantitative arbitrage
Danny’s most pragmatic summary: it can help reduce costs and improve efficiency, but almost impossible to make it open-source for profit.
How serious are security issues?
SlowMist’s Lisa provided the most systematic analysis:
Why doubt OpenClaw’s stability?
Rapid iteration—new version every one or two days, with dozens or hundreds of fixes each update—completely disrupts traditional software engineering rhythm. Under such speed, full testing across devices and scenarios is impossible.
Main risk points:
- Skill poisoning: Malicious plugins can steal account passwords, API keys, tokens, leading to fund theft
- Supply chain attacks: Lobster automatically updates Skills; new versions do not guarantee safety
- Permission abuse: Users with crypto assets on their computers risk fund theft if permissions are misused
Danny’s painful lesson: never let lobster generate wallets and manage private keys, as the returned private key may be fabricated. Skills updates must be manually reviewed; do not let it auto-install.
Teddy’s reminder: when using third-party forwarding, data passes through their servers, risking leakage of API keys and sensitive info. Someone embedded Google API keys, resulting in tens of thousands of dollars in charges.
Principle of least privilege reference:
✅ Can delegate to Agent: coding, document organization, data fetching, information gathering
❌ Must be manually confirmed: involving funds, private keys, core server permissions
When connecting wallets, it’s recommended to use Coinbase Wallet’s Skills, with manual second confirmation for each transfer, and strict isolation.
Are big exchanges’ “skill trees” for lobster reliable?
Binance, OKX have launched related Skills for OpenClaw, but practical traders are generally cautious.
Danny: Only give read-only API to lobster for backtesting; never let it place orders. Less than five orders, okay; more than that, hallucinations will appear.
Todd: The fundamental difference between AI agent trading and quantitative robots is—quant algorithms are trained “AlphaGo,” large language models are just “commentators.” Letting lobster run quant strategies is like having a commentator play professional matches—ineffective.
Teddy: Lobster can be used as an interaction interface, but the underlying logic must be a dedicated Agent trained by you, not just a raw lobster making decisions.
Conclusion: High-frequency quant trading—lobster’s response speed is insufficient; trading decisions—its IQ is insufficient.
Domestic lobster ecosystem vs. overseas, who has more potential?
Danny’s sharpest insight: OpenClaw is essentially a “brain-equipped macro key presser,” very unfriendly to ordinary users, like Linux rather than Windows. Truly skilled users are a tiny minority.
His forecast: in two months, the hype around OpenClaw will fade; the products that truly reach millions will be made by big companies like Tencent, ByteDance, creating “Windows-level” user experiences. The Personal Computer form factor released by Perplexity might be the real mass entry point.
Todd’s observation: domestic enthusiasm is higher because of rapid government intervention (Shenzhen, Wuxi subsidies), and the low cost of domestic models—“gambling cost” is much lower than overseas. Running Claude once costs a few dollars overseas; using Kimi/MiniMax domestically might cost just a few cents, creating a completely different experience.
Opportunities for domestic players?
- Selling courses / installation services: Idle Fish installation earned 260,000 RMB in a few days, but this is an information gap dividend, not sustainable
- Model stocks: MiniMax listed in Hong Kong, rose from 200 HKD to 1,000 HKD—such opportunities are worth noting (not investment advice)
- Crypto payment infrastructure: AI Agents naturally require cross-border, no-KYC, micro-payment settlement methods; USDC micro-payments, native crypto payments are worth continuous attention
- One-person company infrastructure: Lobster makes “one-person company with digital employees” truly feasible for the first time; tools and services around this scenario have great potential
Final advice for all shrimp breeders:
- Manage expectations: Sellers may hype lobster to 150 points, but reality might be only 65
- Use in a simplified manner: Let AI capable of calculus handle addition/subtraction/multiplication/division for best results
- Isolate devices: Play on dedicated machines, separate from work and personal devices
- Fund independence: Any operation involving private keys and funds must be manually confirmed
- Don’t blindly trust Skills: Review before installation, monitor updates, beware supply chain poisoning
- It’s an intern, not a fund manager: Use it for info gathering and decision support, but don’t let it manage funds independently
Note: This summary is based on the record of PANews Space “Shrimp Breeders Alliance: Tencent enters, government subsidies, idle fish installation—how the crypto world responds to “shrimp” anxiety?” Guest opinions do not constitute investment advice.
