
MLA 029 Is OpenClaw Safe? Uses and Alternatives
Machine Learning Guide
Operational Costs and Hosting
Tyler outlines cost ranges, VPS recommendations, and local model options via Ollama to reduce API costs.
OpenClaw in 2026 is safe for read-only jobs on its own machine and accounts, not as a full-access assistant logged in as you. What it's for, how to harden it and vet skills, what it costs, and the alternatives: Hermes Agent, Gemini Spark and OpenAI's dots.
Links- Notes and resources at ocdevel.com/mlg/mla-29
- Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
An open-source, self-hosted personal agent: an always-on agent loop (MLA 28) that you message from a chat app, with access to your files, shell, browser and accounts. Peter Steinberger started it in November 2025. After he joined OpenAI in February 2026, it moved to the OpenClaw Foundation, a US 501(c)(3) that keeps it MIT licensed. The software is free; you pay for the model and the server.
Is OpenClaw safe?Safe enough for read-only jobs on a machine and accounts that belong to the agent; not safe with full access to your own computer and accounts. The 2026 record:
- Tens of thousands of instances exposed to the internet (Censys counted 63,070 live at the end of March).
- A one-click remote code execution bug that stole the gateway token (January) and a pairing-token-to-admin escalation (March); about 720 advisories by November, 14 critical.
- Malicious ClawHub skills: 341 found by Koi Security in February (ClawHavoc), then 575+ reported by Acronis, both delivering stealers through fake "prerequisites".
- 2.0's shared sessions are not a security boundary, secrets aren't encrypted at rest, and sandboxing is off by default.
Injection can't be patched: the prompt-injection docs report adaptive attackers beating defenses over 80% of the time.
OpenClaw security best practices- Where it runs: a dedicated VPS or spare box, gateway on loopback, reached over Tailscale or SSH, auth token always set. In Docker the container binds to the network, and the tool sandbox is opt-in.
- Identity: its own OS user, mailbox, calendar, browser profile and capped API keys.
- Permissions: read-only first; approval on anything irreversible; a tool-less reader agent for email; a strong model for any agent that reads untrusted content.
- Upkeep: openclaw security audit after every change, a pinned version, weekly log reads.
One gateway is one trust boundary; give people who don't fully trust each other separate gateways.
How to vet an OpenClaw skillCheck the ClawHub audit status and risk level → read the whole SKILL.md (reject any step that runs a terminal command or downloads an archive) → read the scripts for unknown network calls and access to env vars, keychain or SSH keys → check the publisher → install into one workspace.
How much does OpenClaw cost to run?The server needs 1 vCPU and 1 to 2 GB of RAM. The model is the bill:
20,000 tokens/turn × 48 heartbeats/day ≈ 1M input tokens/day 1M × $2 per M (Sonnet 5.5 input) ≈ $2/day uncached cache reads at 5% of that price → ≈ $0.10/dayUse a small model for tool-less checks and the strongest model for the agent that acts. The docs advise against local models here (MLA 14 covers self-hosting).
OpenClaw vs Hermes AgentHermes Agent (Nous Research, MIT, Python) writes and improves its own skills, ships with command approval on, and imports an OpenClaw setup. Pick OpenClaw for channels and integrations, Hermes for stricter defaults.
OpenClaw alternativesHosted agents trade control for vendor-run isolation: Gemini Spark (AI Pro and Ultra), OpenAI's dots (Pro and Business Premium), Meta's Muse (free tier, US) and Claude Cowork's cloud scheduled tasks. Tiny self-hosted agents (PicoClaw, ZeroClaw, nanobot) save memory, not tokens.
The full course on this topic is Gnothi's Agentic AI Course, next to Claude Code, at ocdevel.com/courses.
More in the resources. Next: MLA 30, on AI and programming jobs.


