OpenClaw isn't fooling me. I remember MS-DOS
Enthusiasm around OpenClaw-style AI “personal agents” is colliding with deep skepticism over security, reliability, cost, and actual usefulness. Commenters compare the architecture to the unsafe, single-user days of MS‑DOS, warning that giving an LLM broad access to email, files, and online accounts creates huge blast radius for prompt injection, data exfiltration, and accidental damage—especially when sandboxing is coarse and credentials live on the same machine. While some users report genuine productivity gains for coding and narrow automation tasks, the prevailing view is that agent frameworks are still in a toy phase, with unclear everyday use cases and serious safety and privacy questions yet to be solved.
Perceived Value and Real‑World Use Cases
- Many commenters see OpenClaw/agent setups as mostly hobbyist toys or hype; some tried them, found them janky or without a compelling use case, and turned them off.
- Others report concrete value:
- Coding “interns” that implement plans, manage branches, or work on a separate workstation/VPS.
- Infrastructure agents that watch flaky dev servers and auto‑fix issues while logging “learnings.”
- Marketing/social agents that draft posts and graphics, or monitor support queues and nag humans.
- Personal assistants that monitor email (read‑only), calendars, gym schedules, bands’ tour dates, etc.
- Smart‑home control, media downloading, and light IT maintenance.
- Several people compare this to early home computers or 3D printers: currently more about tinkering/learning than net time savings.
Cost, Access, and ROI
- A recurring flashpoint is cost: some users spend around $180/month in API credits; critics call that absurd for “playing music and downloading movies.”
- Comparisons are drawn to live‑in au pairs or cheaper VAs, highlighting how out‑of‑reach this is for “ordinary people.”
- Others argue costs can be cut with cheaper models, local inference, or by using agents sparingly (e.g., Hermes tasks at ~$0.25 each).
- Long‑term GPU economics vs API use are debated; consensus leans toward APIs being cheaper and better for individuals while proprietary models advance rapidly.
Security, Privacy, and Safety Concerns
- Core concern: agents combining private data, untrusted content, and external actions (email, GitHub, payments) are a “ticking bomb.”
- Fear of prompt injection, credential exfiltration, destructive actions (deleting mail, nuking repos) and broad blast radius.
- Many insist they will not give agents payment credentials or high‑risk powers; others cautiously do so with strong limits (read‑only access, manual approvals, separate machines/VPSs, prepaid keys).
Architecture & Sandboxing Debate
- Strong criticism of “sandbox the whole agent” approaches; argued to be MS‑DOS‑like: one big box, no real isolation.
- Advocated alternatives:
- Tool‑level permissions and whitelisted arguments.
- Per‑channel process isolation, encrypted credential vaults, typed secrets, and auditable logs.
- Secret proxies outside the sandbox (HTTP proxies that inject tokens).
- Workflow engines where each new “task” is deployed as a minimal‑privilege app.
- Memory and autonomy are seen as unsolved: naive “heartbeat cron + huge context + RAG” is described as expensive, brittle, and eventually collapsing under its own weight.
Broader Reflections and Historical Analogies
- Comparisons to MS‑DOS, Windows 98, IoT home automation, and “worse is better”: crude, insecure tools that nonetheless may win by being first and convenient.
- Some think OpenClaw‑style assistants are inevitable and will be hardened over time; others see them as a niche for enthusiasts and enterprise “digital glue,” not a mass‑market revolution.