I gave Clawdbot (aka Moltbot) access to my computer, calendar, and emails: Here’s what happened

What It Covers

Claire Vo tests ClaudeBot (MoltBot), an open source autonomous AI agent with computer access, by giving it control of her calendar, email, and desktop. She documents installation challenges, security concerns, workflow experiments including scheduling and research tasks, and evaluates whether autonomous agents are ready for consumer or enterprise use beyond developer experimentation.

Key Questions Answered

• •Installation complexity: ClaudeBot requires two hours of technical setup including installing Homebrew, Node, NPM, Xcode, and configuring dependencies before the single command line installation works. The onboarding process involves creating OAuth clients in Google Cloud Console, downloading JSON credential files, and setting up Telegram bot connections through BotFather, making it inaccessible for non-technical users despite marketing as consumer-ready productivity tool.
• •Security architecture approach: Instead of granting full account access, create a dedicated Google Workspace email for the agent with limited calendar read permissions and a separate 1Password vault containing only agent-specific credentials. Request minimal OAuth scopes for each task rather than accepting default permissions that include delete access across Gmail, Drive, Contacts, and Calendar to prevent unauthorized data exposure or deletion.
• •Time zone calculation failures: ClaudeBot consistently placed calendar events one day late because it attempted to mentally calculate dates rather than trusting API timestamps, requiring manual correction of every recurring event individually since the CLI tool cannot create recurring calendar entries. The agent lacks temporal awareness and struggles with basic date math, making calendar management unreliable without constant supervision and verification.
• •Agent identity and impersonation issues: ClaudeBot defaults to impersonating users rather than identifying as an assistant, sending emails signed with the user's name from a different email address. Explicit prompting is required to make the agent identify itself as an assistant in all communications, and even then, the bias toward acting as the user rather than on behalf of the user persists throughout interactions.
• •Asynchronous research workflow success: Voice-initiated research tasks where ClaudeBot analyzes Reddit threads for product feedback and delivers markdown reports via email provides the optimal use case. The latency becomes acceptable when tasks genuinely require research time, the multi-channel communication (voice input, email output) mirrors working with human employees, and the structured output format enables immediate roadmap decisions without additional processing.