Then April hit. And everything fell apart.
The Astroturfing Question: Real Growth or Manufactured Hype?
Steinberger did end up at OpenAI shortly after OpenClaw peaked. The timeline is suspicious enough that even his defenders admit the optics are terrible.
But the “it was all bots” narrative doesn’t hold up under scrutiny. The project’s GitHub activity tells a different story. In the past week alone, OpenClaw had 2,700 merged pull requests (1,000 opened), 3,000 commits from 266 contributors. Compare that to llama.cpp, which had 80 PRs merged and 116 commits from 51 authors in the same period.
That’s not the signature of a dead project. That’s the signature of a project that’s still being actively worked on by a massive community, even if the mainstream conversation has moved on.
What Actually Killed the Hype?
1. The Usage-Based Pricing Reality Check
OpenClaw is free software. The operational costs, however, are anything but. Light users might spend $5, $20/month on LLM API calls, but heavy users regularly blow past $100/month. When you factor in always-on background agents running heartbeat schedulers against premium models, costs can hit $300+/month.
Here’s the breakdown that nobody talks about:
| Model Tier | Price per 1M Input Tokens | Monthly Cost (Heavy Usage) |
|---|---|---|
| Google Gemini Flash | ~$0.05, $0.10 | Minimal |
| DeepSeek Models | ~$0.20, $0.30 | Low |
| Anthropic Claude (Sonnet) | ~$0.50, $3.00 | $50, $150 |
| OpenAI GPT-4-class | ~$5, $15+ | $100, $300+ |
| Local Models (Ollama) | $0 (API cost) | Hardware cost only |
The killer feature that nobody warned about: the heartbeat scheduler. Even when you’re not actively using OpenClaw, it’s polling your system, running background checks, and burning tokens. Developers who thought they were getting a free assistant discovered they’d signed up for a token-metered subscription with no ceiling.
2. The Stability Meltdown (April, June 2026)
This is where the technical story gets interesting. Between April and June 2026, OpenClaw went through a rough patch where almost every release broke something. Users reported that skills that worked perfectly in March would fail silently in April. The WhatsApp integration, supposedly a flagship feature, was notoriously fragile and required frequent re-authentication.
The situation degraded to the point where, as one user put it, “It was a case study of a purely vibe-coded product.”
This coincided exactly with the period when Hermes (from Nous Research) was hitting its stride. Hermes was stable, had more features, and most critically, offered a migration path from OpenClaw that could auto-import settings, memories, skills, and API keys. The door didn’t just open, it had a welcome mat.
3. The Conway’s Law Problem
OpenClaw’s architecture was designed around a specific vision: a monolithic gateway that handled everything from messaging to skill execution to memory management. It worked beautifully when Steinberger was the sole maintainer. But as the community grew and the codebase expanded, the architecture couldn’t keep up.
The 2026.3.31 release introduced “Task Brain”, a SQLite-backed task ledger that consolidated agent control protocol tasks, subagents, cron jobs, and background processes into a single management layer. This was described as “Kubernetes for agent tasks.” It was also a recognition that the original architecture had become unmanageable.
By contrast, Hermes took a more modular approach. Its skill system was cleaner, its memory system more transparent, and its deployment path significantly simpler. For most users, the choice wasn’t between equal contenders, it was between a project still fighting its own complexity and a project that had already learned those lessons.
Was OpenClaw Actually Useful? The Real Use Cases
Daily Triage and News Monitoring: Several users report successful long-term use for automated news aggregation and daily briefings. One user noted their OpenClaw still handles “daily triage of news” months after the hype died.
DevOps and System Monitoring: The persistent background agent model worked well for monitoring server logs, deployment pipelines, and uptime metrics. Configured correctly, it could restart services and create incident tickets without human intervention.
Local-First Privacy: When running with local models via Ollama, OpenClaw offered genuine privacy advantages. Legal professionals processing confidential contracts found value in an agent that never touched third-party servers.
The Skeptics’ Counter: A significant number of users found the tool actively unhelpful. One expert reported that when they used OpenClaw for research in areas they knew well, “half of the stuff it came up with was just nonsense or didn’t exist.” The memory system, in particular, was criticized for accumulating falsehoods that the agent would then treat as facts, a feedback loop that degraded performance over time.
The Deeper Cost Crisis
This context makes OpenClaw’s usage-based pricing model look less like a greedy decision and more like the only viable option. The alternative, flat-rate pricing, would have required either unsustainable subsidies or pricing out the very community that made the project popular in the first place.
The Hermes Factor: Why the Competitor Won
Hermes didn’t try to be everything to everyone. It focused on a narrower set of capabilities and executed them exceptionally well. Its migration tooling, which could import OpenClaw settings with zero friction, was a masterstroke. The message to OpenClaw users was clear: You don’t have to start over. Just move over.
The detailed comparison between the two frameworks reveals fundamental philosophical differences:
- OpenClaw: Workflow-driven, excels at tool orchestration and stable execution. Best for standardized automation tasks.
- Hermes: Learning-driven, emphasizes experience retention and capability optimization. Better for complex, long-term tasks.
For most developers, Hermes’s approach won because it required less upfront configuration and degraded more gracefully. When OpenClaw broke, it broke hard. When Hermes struggled, it usually still worked, just less efficiently.
What the Numbers Actually Say
- 2,700 merged PRs in the past week (vs. 80 for llama.cpp)
- 3,000 commits from 266 contributors in the same period
- 68,000+ GitHub stars (post-hype, down from the 196,000 peak during the Clawdbot era)
- 50,000+ active users according to the project’s own metrics
These aren’t the numbers of a dead project. They’re the numbers of a project that went from mainstream hype to niche utility. The developer community still exists, it’s just not loud anymore.
Is There Still a Future for OpenClaw?
But the window for mainstream adoption has likely closed. The AI agent space moves too fast for second chances. By the time OpenClaw sorted out its stability issues, Hermes had already captured the community’s attention. By the time the pricing concerns were addressed, the discourse had moved on.
The broader examination of whether OpenClaw’s virality was organic or manufactured reveals that the truth is probably somewhere in between. There was real enthusiasm from developers who genuinely wanted an open-source autonomous agent. But that enthusiasm was amplified, possibly artificially, and when the amplification stopped, the crash was inevitable.
Takeaways for the AI Community
- Usage-based pricing kills adoption. OpenClaw’s death wasn’t caused by bad code, it was caused by a pricing model that turned a free tool into a money pit once background agents started running. If you’re building on top of LLM APIs, make the costs transparent before users commit.
- Stability beats features. During the April, June meltdown, OpenClaw was pushing new features every week. What users wanted was for existing features to work reliably. The project prioritized velocity over reliability, and it paid the price.
- Migration paths matter. Hermes’s ability to import OpenClaw settings was a killing blow. When a competitor makes it this easy to leave, you’d better make sure your users never want to leave.
- Hype is a liability. Projects that rocket to 196,000 GitHub stars in months create expectations that can’t be sustained. The crash is inevitably harder than it would have been with organic growth.
The Bottom Line
The code is still there. The community still exists. And for users who need maximum control over their agent infrastructure and are willing to invest the time to configure it properly, OpenClaw remains a viable option. But the hype window has closed, and it’s not coming back.
The deep dive into OpenClaw’s token-burning economics shows that even the project’s most enthusiastic supporters eventually hit the same wall: the cost of running an always-on agent against premium models is unsustainable for most individual developers.
In the end, OpenClaw’s story isn’t about marketing manipulation or technical failure. It’s about a project that tried to do too much too fast, and got punished by a market that values stability over speed. The tools that survive this hype cycle won’t be the flashiest. They’ll be the ones that work reliably, cost predictably, and don’t break every Tuesday.
That’s not a conspiracy. That’s just engineering.
