One week with openclaw

OpenClaw is far from being a consumer-ready AI assistant or a generalist agent. After spending a week trying to set it up, two issues stand out above all others: cost and configuration.

Agents and sub-agents

Long-running sessions are expensive. As context accumulates, token usage grows, and much of that context simply isn’t needed anymore. Humans are remarkably good at discarding irrelevant details and remembering only what matters at a given moment. We subconsciously separate signal from noise. LLMs, on the other hand, require a much more deterministic approach.

With OpenClaw, you have to explicitly decide what gets remembered¹, what gets passed along, and what gets dropped. This is essential not just for better results, but also for keeping costs under control. Without that discipline, conversations become bloated, inefficient, and unnecessarily expensive.

Configuration: OpenClaw cannot configure itself

Running OpenClaw comes in two main flavors: run local or run on a server.

You shouldn’t be running openclaw on a personal laptop, especially for users who aren’t deeply familiar with the underlying technology. A server—either self-hosted or via a VPS—is a more realistic option, but requires even more technical knowledge to implement. You need to understand infrastructure, deployment, security, and ongoing maintenance. On top of that, you have to factor in the cost of the VPS itself.

Configuration is where things really break down. This isn’t an AI to which you can address saying – “Configure yourself” – and expect usable results. I have a fairly broad understanding of the systems OpenClaw is built on, and I still tried to configure it by talking to OpenClaw directly. Within a couple of days, I had burned through nearly $100—and not because it was working well.

Eventually, I pivoted to a more controlled approach: writing code directly, or using tools like Cursor. While this wasn’t free either, it gave me far more control. Single-threaded chat-based configuration inside OpenClaw simply didn’t produce the results you’d expect for the cost involved.

A concrete example: after polishing most of the setup, I asked OpenClaw to add some guardrails. The goal was simple—avoid large, expensive outputs and check in with me more frequently before producing long responses. OpenClaw suggested creating a new file called GUARDRAILS.md.

The problem?: “you’re absolutely right!” – as the agent replied to my follow up question – the file isn’t automatically included in the prompt when a new session starts. In other words, the guardrails don’t actually apply. This is a basic failure mode, and fixing it requires a deeper understanding of how prompts are constructed and injected. Your best chance of getting a correct interpretation of what you want is by using the most expensive model. And even then, initial responses may be flawed.

Crafting a solid system prompt for your specific use case is critical, and you should not rely on OpenClaw itself to do that work for you.

Cost

Cost is one of the biggest factors in any LLM-based setup.

If you want a personal AI today, the simplest option is paying subscriptions. Most services charge somewhere between $15 and$20 per month. If you rely on multiple providers—say OpenAI, Anthropic, and another vendor—you’re quickly looking at around $50 per month.

The downside of subscriptions is flexibility. You don’t get the same level of control, personalization, or advanced reasoning workflows that OpenClaw promises out of the box.

With OpenClaw, however, the cost problem takes a different shape. The cheaper models simply don’t deliver acceptable results. I experimented with Claude (both Opus and Sonnet 4.5) and with GPT models, including GPT-5.1 (thinking) and GPT-5.2. The higher-tier Claude models were extremely expensive, so I fell back to the cheaper GPT options to test performance.

That turned out to be a mistake. For several days, I got poor results while simultaneously trying to configure large parts of OpenClaw. Because I wasn’t using the system in a stable, “normal” way, it became difficult to even evaluate model quality. The combination of high configuration overhead and inconsistent output made the entire process inefficient and frustrating.

Conclusion

In its current form, OpenClaw is a powerful but demanding tool. It rewards deep technical knowledge and careful cost management, but it’s nowhere near ready for mainstream users who just want an assistant that works.

Advanced workflows are the ones where you will really make a good ROI: Autonomous conversations or multi-agent workflows are two examples, where you can immediately see a a high return. The configuration required for these, and the price you’d pay, make for a difficult choice as agents make their way to the tools you work daily. Most common use cases ranging from low to medium complexity can already be handled by those apps: Notion AI, Gemini y the Google Suite, Slack AI, etc.

Footnotes

1. You don’t have to, OpenClaw will do this for you, but won’t be very smart about it without spending 5$ per message. ↩