OpenClaw review: costs, security and setup limits

OpenClaw connects messaging channels to an AI agent that can call tools and retain conversation history. This review explains its gateway, ReAct runtime and scheduled tasks, then examines the maintenance and trust problems the speaker says drove some early users away.

The architecture section follows a message through conversation routing, memory retrieval and repeated model and tool calls. Built-in cron can trigger the same process for a scheduled email summary. Running this self-hosted software does not eliminate model API costs: the speaker cites a user's estimate of roughly $86 per month for heartbeat activity under their default configuration, rather than a fixed OpenClaw fee.

Reported problems include silent OAuth failures and lost memory after an upgrade to version 2026.03.2. The speaker also recounts Summer Yue's inbox deletion incident, attributing it to context compaction dropping a confirmation rule. Security examples include an email prompt injection that allegedly exposed an SSH key and a localhost trust setting combined with a misconfigured reverse proxy.

The practical advice is to automate one narrow workflow, reserve expensive models for demanding tasks and isolate the agent in a virtual machine or sandbox. These are the speaker's recommendations for limiting cost and access, not guarantees of reliable or secure operation.