Clankers with claws
TLDR
- DHH ran OpenClaw on an isolated Proxmox VM, gave it no MCPs or APIs, and watched it autonomously sign up for HEY email, Fizzy, and Basecamp without a single correction.
Key Takeaways
- OpenClaw gives an AI model persistent execution, long-term memory, email access, and reminders outside the prompt-response cycle.
- With only a natural-language prompt and an invite link, the agent named Kef signed up for a HEY email account and two collaboration tools in sequence, zero errors.
- The same experiment was repeated on Kimi K2.5 via OpenRouter with comparable results, though noticeably slower than Claude Opus 4.5.
- DHH deliberately installed no skills, MCPs, CLIs, or API integrations – the agent navigated human-facing web UIs cold.
- The tradeoff is speed and token cost; DHH frames MCPs and CLIs as a temporary crutch, not a permanent requirement.
Why It Matters
- An agent that works with standard human web interfaces needs no custom tooling per service, which removes the integration bottleneck for deploying personal agents.
- Running the agent on an isolated Proxmox VM with no access to personal data is the minimum safe setup; DHH notes others are running it on their own machines despite explicit risk warnings.
- If agents stop requiring LIDAR-style accommodations (MCPs, custom APIs), the cost of standing up new agent workflows drops to writing a prompt.
· 2026-02-05 · Read the original
