My 24 Hours with Clawdbot (aka Moltbot)—3 Workflows for a Powerful (and Terrifying) AI Agent

By: Claire Vo Source: How I AI (ChatPRD / Lenny’s Podcast Network) Type: podcast

Summary

Hands-on 24-hour test of Clawdbot, an open-source autonomous agent, with unfiltered assessment. Setup (Workflow 1): ~2 hours of dependency installation (Homebrew, Node, Xcode CLI tools) on a dedicated old MacBook with isolated user account. Connected via Telegram (BotFather API token). Created completely separate digital identity — new Google Workspace email, dedicated 1Password vault. Chose Sonnet 4.5 for safety/cost reasons over Opus. Calendar EA (Workflow 2): Simple single-event scheduling worked via workaround (agent creates on its own calendar, invites you) — but initial OAuth requested all Google scopes when only calendar-read was needed (agent admitted over-requesting when challenged). Complex family calendar scheduling failed badly — every event off by one day, can’t create recurring events, and bot tried to re-add events while Claire was deleting them. Async work (Workflow 3): Coding a Next.js conversation viewer worked locally but latency too slow for coding feedback loops. Reddit market research was the star — voice command via Telegram produced excellent structured Markdown report with actionable insights, delivered to email. Verdict: Uninstalled after 24 hours due to security concerns. Has category-level product-market fit but isn’t ready for non-developers.

Key Ideas Extracted

Question every permission scope: Agent initially requested all Google scopes (files, contacts, calendar, email) for a calendar-read task — immediately backed down when challenged. Always push back on excessive permissions.
Act FOR me, not AS me: Constant tension between agent impersonating the user vs. working on their behalf — this is a fundamental design challenge for autonomous agents
Dedicated machine + isolated accounts: Security best practice — separate user account, dedicated email, separate password vault with only agent-specific credentials
LLMs can’t reliably handle time/dates: Calendar events were consistently off by one day; agent admitted to “mentally calculating” dates instead of using API data — fundamental limitation
Research > real-time tasks for agents: Async research (Reddit scrape → synthesized report) is ideal because latency is a feature, not a bug — matches natural delegation patterns
Voice → action → email delivery: Compelling UX pattern — give voice command via messaging app, receive polished deliverable in inbox
Category has PMF, product doesn’t: Clawdbot proved the concept of always-on autonomous agents but the specific implementation is too risky/technical — open question of whether big tech or startups will build the viable version
Workaround pattern for trust: Instead of granting write access to your calendar, have the agent use its own calendar and invite you — pattern applicable to other permission-sensitive domains

Notes

Published Jan 28, 2026 on ChatPRD blog. 11-min read. Solo episode.
Named the Clawdbot instance “Polly”
Chose Sonnet 4.5 specifically out of fear of what Opus might do autonomously — interesting model selection rationale
Tools: Clawdbot/Moltbot (open-source agent), Telegram (messaging interface), Google Workspace (email/calendar), 1Password (credential isolation), Anthropic API
The family calendar disaster is a cautionary tale — deleted events being re-added by the bot while she was on the phone at Target
Ultimately uninstalled and deleted all keys/tokens after the test
Sponsor: Lovable

Raw Content

Re-scraped from ChatPRD 2026-02-15. Full article content captured in Summary and Key Ideas above.