Is OpenClaw, billed as an "AI worker", worth using?

Short answer: not yet

OpenClaw has been hyped as an "AI worker" that can “take over your computer and free your hands,” but it has been reported that the tool functions more as a coordinator than as a brain: it interprets user instructions, calls out to external large models and orchestrates workflows, while the heavy lifting is done by the connected LLM. So is the promise real? The short answer from a recent deep test by 每日经济新闻 is: not yet.

What the tests found

Reporters wired OpenClaw to a roster of Chinese and Western models — including Qwen3-Max (千问 Qwen3-Max), Kimi-K2.5 (月之暗面 Kimi-K2.5), MiniMax-M2.1/M2.5 (MiniMax), GLM-4.7 (智谱 GLM-4.7) and GPT-5-mini (OpenAI) — and asked it to find a local interview transcript, supplement with web research, draft a profile piece and email the稿件. Results varied wildly. Qwen3-Max struggled to locate files and repeatedly stalled on sending mail; Kimi could read a file and summarize it but hit 429 web errors during browsing and failed the mail step; MiniMax-M2.1 handled retrieval and drafting but only sent a fragment when asked to email, while MiniMax-M2.5 completed the whole chain; GLM-4.7 typed an incorrect mail URL and needed correction; GPT-5-mini was the most consistent, completing almost the full workflow with minimal intervention. Common failure modes were browser-control errors, wrong DOM/input-node handling, repeated instruction loops and network-rate limits.

Expert read: a "task framework" that depends on the model

Industry voices say OpenClaw behaves like a task framework whose output quality is determined by the underlying model and by brittle browser-automation glue. It has been reported that some observers compare OpenClaw to a packaged version of developer tools such as Anthropic’s Claude Code — a friendlier shell but not a new foundational capability. Experts warn of practical hurdles: technical setup, cost (accessing leading foreign models can be expensive), and security risks from automated access to local files and web sessions. Geopolitics matters too: Chinese firms are racing to close capability gaps under the weight of export controls and data-security concerns, but for now Western models still often offer higher ceilings — if you can afford or access them.

Bottom line

OpenClaw is interesting for technophiles and as a preview of agent-style automation, but the deep test suggests it is far from a plug‑and‑play productivity tool for most users. Will an “AI worker” replace your assistant today? Not unless the underlying models and browser-integration become a lot more robust — and unless companies can demonstrate safer, cheaper, and more reliable operation.