OpenClaw: 325K GitHub stars in 4 months.
Monday at GTC, Jensen Huang: "Every company needs an OpenClaw strategy." Tuesday, Anthropic ships Dispatch and Claude Code Channels.
-
Dispatch: text Claude from your phone, it runs tasks on your desktop via Cowork. Sandboxed. Local-first. Asks before touching anything sensitive. From $20/mo: https://support.claude.com/en/articles/13947068-assign-tasks-to-claude-from-anywhere-in-cowork
-
Claude Code also shipped Channels: push tasks from Telegram or Discord into a running session: https://code.claude.com/docs/en/channels
First impression: Dispatch made less noise than the remote-control command a few weeks back.
But I see the same question answered two very different ways: what should an AI agent be allowed to do without asking?
-
OpenClaw: configure everything, trust the user.
-
Anthropic: sandbox everything, confirm before acting.
Both break at scale.
Configuring 100+ skills is a security surface nobody can audit. Cisco already caught a third-party OpenClaw skill exfiltrating data and injecting prompts without the user knowing.
On the other side, confirming every file move, every email draft, every calendar update defeats the purpose of delegation. Early testers report Dispatch succeeds about half the time on complex tasks.
The company that nails the autonomy part, not "trust everything" or "confirm everything", wins this market.