I Let an AI Agent Auto-Trade Crypto for 2 Weeks. OpenClaw Is Brilliant - and Terrifying.

If you’ve been anywhere near tech Twitter this month, you’ve seen it.

Screenshots of six-figure profits. Photos of Mac Minis lined up on desks, captioned “my new employees.” Threads promising passive income powered by an autonomous AI lobster.

Welcome to the OpenClaw era. And I went all in.

I’m a developer turned day trader. I write code, I read charts, and I have a fairly high tolerance for things that break. So when I saw that OpenClaw — the open-source AI agent that racked up 150,000+ GitHub stars in weeks — could be connected to live trading systems, I didn’t just read about it.

I built a setup and ran it with real money.

Here’s what happened, what I learned, and why I think most people getting into this are dangerously underprepared.

What OpenClaw Actually Is (And What It Isn’t

Let me clear something up first, because the marketing around this has gotten out of hand.

OpenClaw is not a trading bot. It’s a general-purpose autonomous AI agent that runs locally on your hardware. It can talk to you through Telegram, control files and scripts on your machine, call APIs, and execute multi-step workflows based on natural language instructions.

The trading part comes from community-built “skills” — essentially plugins that connect the agent to exchanges, wallets, and data feeds.

The project started as “Clawdbot” by Peter Steinberger, got hit with a trademark complaint from Anthropic, renamed to “Moltbot,” had its social handles sniped by scammers who launched a fake token that pumped to $16M and crashed 90%, renamed again to “OpenClaw,” and then went properly viral. Three names in two months. It’s chaos. Beautiful, terrifying chaos.

My Setup: OpenClaw + OpenAlgo + Crypto Futures

I ran OpenClaw on a Mac Mini M4 with 16GB RAM. For the trading layer, I used OpenAlgo — an open-source platform that provides a unified API across multiple brokers. The stack:

Data: Live price feeds via exchange WebSocket APIs
Strategy logic: A momentum-based approach I’d been running manually, translated into natural language instructions for the agent
Execution: OpenAlgo API for order placement and position management
Risk rules: Hard-coded stop losses outside the agent, plus a daily loss limit kill switch

The whole point was to see whether an AI agent could handle the mechanical side of trading — monitoring setups, entering positions, managing exits — while I focused on higher-level strategy decisions.

Setup took about an hour. That includes installing OpenClaw, configuring the OpenAlgo bridge, writing the initial skill, and connecting my exchange API keys through the broker’s official system. Not through OpenClaw’s own key management — that’s a security line I won’t cross.

Week 1: The Honeymoon

Honestly? It was impressive.

The agent identified three setups on day one that I would have taken manually. The entries were clean. It sized positions according to the rules I gave it. When one trade hit my trailing stop, it exited without hesitation.

By the end of week 1, the account was up roughly 8%. Not spectacular, but not the point. What mattered was that it executed my rules better than I do. No FOMO entries. No “let me hold this a bit longer” on a position that’s clearly done. No revenge trades after a loss.

The bot doesn’t have emotions. And in trading, that’s worth more than alpha.

I also set up a monitoring workflow where the agent summarised overnight market moves via Telegram every morning — pulling data from multiple sources, identifying key levels, and flagging anything unusual. That alone saved me an hour a day.

Week 2: The Cracks

Then things got interesting.

On day 9, the agent took a trade that made no sense. It entered a long position on ETH during a clear downtrend, against every rule I’d specified. When I reviewed the logs, I found that a skill update had subtly changed ㄴhow the agent interpreted one of my conditions. The natural language instruction “enter long when momentum confirms above the 20 EMA” was being parsed differently after the update.

This is the core problem with AI-driven auto-trading: **your strategy is mediated through language, not code.** In traditional algo trading, a condition is either true or false. With an LLM-based agent, a condition is interpreted — and interpretations can drift.

I caught it because I was monitoring. Most people following the “set it and forget it” narrative won’t.

Day 12 brought a scarier moment. I’d installed a third-party skill from ClawHub to enhance the agent’s market analysis. Three days later, I ran a security audit and found the skill was making outbound API calls I hadn’t authorised. Nothing malicious in my case — it was sending anonymised usage data — but it could have been anything.

This isn’t hypothetical paranoia. Kaspersky found 512 vulnerabilities in OpenClaw’s codebase, eight of them critical. 386 malicious skills have been discovered on ClawHub.

Gartner called it
A dangerous preview of agentic AI, demonstrating high utility coupled with unacceptable cybersecurity risk.

I removed the skill, revoked the API keys, and regenerated new ones through my broker.

The Numbers

After 14 days, here’s where I landed:

- Net PnL: +4.8%
- Total trades: 47
- Win rate: 57%
- Max drawdown: -6.2%
- Trades I manually overrode: 4 (including the rogue ETH long)

Decent return. But let me be honest — I could have achieved similar results running the strategy manually with less risk. The value wasn’t in the PnL. It was in the time saved and the emotional discipline the agent enforced.

The Real Problem With AI Auto-Trading

Here’s what nobody talks about in those viral threads.

**The edge doesn’t come from the tool.** OpenClaw is infrastructure. It executes instructions. It doesn’t generate alpha. The people posting six-figure screenshots either had a genuine edge before OpenClaw (speed arbitrage, prediction market inefficiencies, privileged data) or they’re posting fabricated results for engagement.

A developer recently shared a workflow using OpenClaw + Microsoft’s Qlib framework that backtested at 59% annualised return. The technical stack was legitimate — Qlib is a serious quant research platform. But as Qlib’s own documentation states, backtest returns and live performance are structurally different. Transaction costs, slippage, liquidity constraints, and overfitting risk compress real returns dramatically. That 59% is a hypothesis, not a result.

**Security is genuinely concerning.** You’re giving an autonomous agent access to your exchange accounts, your messaging apps, and your local filesystem. One compromised skill, one prompt injection attack embedded in a market data feed, and your entire setup is exposed. This isn’t theoretical — it’s already happened to people.

**Regulation is coming.** The CFTC has already published rules around automated trading systems. Right now, crypto prediction markets exist in a grey area. That won’t last forever. Building your income around a setup that might become non-compliant is a time bomb.

Would I Keep Using It?

Yes — but not for auto-trading.

I’ve stripped back OpenClaw’s permissions. It no longer executes trades directly. Instead, it acts as a research assistant and signal generator: monitoring markets, aggregating data, summarising sentiment, and flagging setups for me to review and execute manually.

That middle ground — AI-assisted trading rather than AI auto-trading — is where I think the real value is for individual traders right now. You get the analytical horsepower and the time savings without the security exposure and the risk of an agent misinterpreting your intent.

Fully autonomous AI trading will happen. The infrastructure is improving fast, the open-source ecosystem is maturing, and projects like OpenAlgo are building the execution layer properly. But we’re not there yet. The gap between “technically possible” and “reliably safe with real money” is still enormous.

If you’re a developer who understands the stack, can audit your own skills, and has proper risk controls outside the agent — there’s something genuinely powerful here. If you saw a Twitter thread and want to set up a Mac Mini that prints money while you sleep — please don’t.

The lobster is smart. But it doesn’t care about your money.

If you found this useful, follow me. Next up: I’m testing whether OpenClaw can replace my entire pre-market research workflow. No trading, just analysis. That might actually be the killer use case.