OpenClaw
- Entity ID:
ent-20260423-p3a000000001 - Type:
project - Scope:
private - Status:
active
Description
An independent open-source AI agent system (Steinberger and OpenClaw Contributors, 2026) described in arXiv:2604.14228 as a local-first WebSocket gateway (default port 18789, loopback-only) connecting roughly two dozen messaging surfaces (WhatsApp, Telegram, Slack, Discord, Signal, etc.) to an embedded agent runtime with companion apps on macOS, iOS, and Android. Used as the paper's primary comparative case study to calibrate Claude Code's design decisions by showing how the same recurring design questions yield different architectural answers in a multi-channel personal-assistant deployment context.
Key claims
- Same design questions, different answers with deployment context
- Claude Code vs OpenClaw: per-action vs perimeter safety
- Claude Code vs OpenClaw: loop-as-center vs loop-as-component
- Claude Code vs OpenClaw: per-agent context vs gateway capability surface
- Multi-agent architecture is the starkest divergence
- Two systems are compositional, not exclusive alternatives
- Opposite bets follow from different trust models and deployment topologies
- Memory investment priorities diverge
Relations
- OpenClaw --[contains]--> Perimeter-Level Access Control (OpenClaw)
- OpenClaw --[contains]--> Gateway Control Plane Architecture
- OpenClaw --[contains]--> Manifest-First Plugin System (OpenClaw)
- OpenClaw --[contains]--> OpenClaw Multi-Agent Routing
- OpenClaw --[contains]--> OpenClaw Workspace Bootstrap File Set
- OpenClaw --[contains]--> Experimental Dreaming System (OpenClaw)
- OpenClaw --[contains]--> ACP (Agent Client Protocol) Composition
- ACP (Agent Client Protocol) Composition --[contains]--> OpenClaw
- Recurring Design Questions Framework --[compared_with]--> OpenClaw