Open Claw

Overview

OpenClaw is an open, local-first personal-agent product or framework used across the sources as a marker of the shift from chatbots to agents that remember context, use skills and tools, schedule work, interact through messaging surfaces, and act on local or external systems. Its importance in the wiki is less a unique model breakthrough than a packaging and interaction breakthrough that made persistent, permissioned agents legible to a wider audience.

Current Profile

OpenClaw demonstrates a personal AI computer pattern: model capability is wrapped in memory, skills, resources, scheduling, input/output channels, APIs, and an agent loop. That form can automate research, office work, reminders, content processing, and coding-adjacent tasks, and it can serve as a probe for workflows that later become engineered products. Its present value is concentrated where a user already has a worthwhile workflow and can supervise it; without that demand, setup, token spend, and risk can dominate. The same access that makes it useful creates its central weakness—unstable memory, configuration, code quality, compounding execution cost, prompt injection, uncontrolled actions, unclear identity or permission boundaries, and a community-skill supply chain whose packages can exercise real authority.

Key Characteristics

• Packages model reasoning with persistent memory, skills, tools, schedules, channels, and local execution.
• Uses messaging and always-on interaction to make agent capability accessible beyond command-line specialists.
• Functions as a programmable action layer across personal, office, research, and operational workflows.
• Encourages user-specific training through feedback, context files, standards, and recurring routines.
• Exposes security, reliability, cost, configuration, and permission failures because it can touch real files, accounts, tools, and external systems.
• Makes skill provenance and package review a security concern because community extensions can inherit the agent’s permissions.
• Serves as a category signal and prototyping substrate rather than a proven final form for personal or enterprise agents.

Evidence

Product mechanics and interaction breakthrough

• 20 个问题，搞懂 OpenClaw：爆红机制、本质变化、创业机会 attributes OpenClaw’s impact to messaging interfaces, local execution, memory, skills, tools, and feedback loops rather than a uniquely stronger base model.
• 139. 【Agent的综述】和苏煜聊Agent技术史、OpenClaw Moment、边界的消弭和社会的辐射 calls the shift an OpenClaw moment in which permissions, always-on operation, personal context, and reachability changed public understanding of agents.
• Jensen Huang LIVE: Nvidia’s Future, Physical AI, Rise of the Agent, Inference Explosion, AI PR Crisis records Huang’s description of memory, skills, resources, scheduling, I/O, and APIs as the components of a personal artificial-intelligence computer.

Personal workflow and local infrastructure

• Vol. 161 从开发自己的 OpenClaw 聊起 treats tools, channels, skills, triggers, and permissions as the agent’s product surface and describes isolation in a virtual machine.
• Vol. 167 Token 如流水，Agent 似朝阳 uses separate messaging sessions for topic-specific settings, memories, permissions, collection, translation, calendar, notes, and task workflows.
• 这半年，我们又买了哪些科技好物？ links always-on local agents to Mac mini and reused-device infrastructure.
• OpenClaw 之后，我只想未来 3-6 个月的事情｜对谈 Sheet0 创始人王文锋 interprets coding agents as a general action layer extending beyond software engineering.

Reliability, safety, and cost

• 当可靠的代码变成了偶尔发疯的OpenClaw，我们未来的工作范式变迁 reports costly model calls, routing, task failure, configuration mutation, prompt injection, and continued action as reasons to keep high-impact work bounded and reviewable.
• E249｜Token经济转点：OpenClaw、Hermes到本地自研的Agent进化之路 credits the personal-agent threshold while criticizing code quality, stability, setup, memory, and token efficiency.
• E163.要完了？不！是要玩了！论养AI的心态与习惯 frames useful “raising” of an agent through feedback and quality gates while warning against an endless attention loop.
• No.219 关于 OpenClaw，到底是谁养了虾，虾又会养谁？ argues that autonomous loops can compound an early mistake into large token waste and real-world harm, especially under broad local permissions or untrusted skills and web content.
• EP 30: OpenClaw: The Open-Source AI Agent That Got Its Creator Hired by OpenAI adds a reported third-party-skill exfiltration and prompt-injection case, corporate-ban claims, and advice to keep current deployments sandboxed and off the user’s primary machine.

Workflow fit and delegated interaction

• No.219 关于 OpenClaw，到底是谁养了虾，虾又会养谁？ locates near-term value among users with existing development, content, investment, coordination, or small-business workflows and warns that tool-first side-business hopes rarely create their own demand.
• No.219 关于 OpenClaw，到底是谁养了虾，虾又会养谁？ frames OpenClaw as evidence for Delegated Agent Interaction: users hand over outcomes rather than request answers, while people retain goals, audit, creative judgment, and responsibility.

Enterprise and ecosystem interpretations

• Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ treats OpenClaw’s attention as an industry signal without claiming the mature software form is settled.
• 为什么公司用不好AI？从焦虑到行动的 3 个关键动作｜对谈百融智能张韶峰 shows how it made agent-driven office workflows imaginable to traditional businesses while leaving adoption paths unclear.
• 我们是如何定义 OpenClaw for Teams 新产品形态的｜对谈 Kuse&Junior 联创兼 CTO 宇豪 says team use requires company memory, identity, authority, security evaluation, auditability, and enterprise-specific economics.
• 138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 treats the framework as a middle layer for context, memory, tools, cost routing, agent data, skills, and post-training.
• OpenClaw 之后，谁将定义主动式 AI 的新战场？｜对谈 AirJelly 黄柏特 argues that execution alone may be easier to commoditize than intent, operating-system context, and durable personal memory.
• Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维 shows nontechnical users recognizing value when the agent programmatically fetches, processes, and pushes work.
• 当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent places memory, skills, and harness reliability at the center of the broader domestic agent wave.

Qualifications

• Bytes: Week in Review - Amazon and AI, YouTube tops the media market and Meta buys an AI-only social network creates an unresolved identity ambiguity by describing an OpenClaw parent/founder relationship around MoteBook and OpenAI; it is not merged here as settled corporate history.
• EP 30: OpenClaw: The Open-Source AI Agent That Got Its Creator Hired by OpenAI attributes the project to Peter Steinberger and supplies another rename, growth, hiring, and foundation account. This partly clarifies the earlier MoteBook ambiguity but conflicts with other creator and origin descriptions, so the biography remains source-scoped.
• The sources mix product observation, practitioner experimentation, forecasts, and analogies; they do not establish stable benchmark superiority or mass enterprise adoption.
• Persistent local access increases privacy and control for some users while also increasing blast radius if permissions, credentials, or prompts are mishandled.
• Always-on usefulness can generate substantial token, hardware, setup, supervision, and maintenance costs.
• Personal-agent memory and permissions do not transfer safely to team use without explicit enterprise identity, audit, and authority models.
• The No.219 source’s adoption, cost, security-incident, project-history, ranking, and corporate-action details are host-reported rather than independently documented in the transcript.
• EP30’s GitHub counts, company bans, security finding, autonomous dating-profile anecdote, trademark chronology, and OpenAI employment claim are likewise host-reported; “Mold Block” and “Mold Match” may be naming errors.

What Changed

• Added community-skill provenance and review as a distinct part of OpenClaw’s security boundary.
• Strengthened the case for sandboxing and withholding main-machine access during current experimentation.
• Added Peter Steinberger’s source-attributed creator and OpenAI trajectory without resolving conflicting origin accounts.
• Preserved the judgment that OpenClaw may be historically important without being the mature final form.

Relationships

• Agentic Software - broader software architecture that OpenClaw helped make visible.
• Persistent Agent Memory - continuity layer central to personalization and long-running work.
• AI Skills - packaged procedures and tools through which the agent gains capabilities.
• Agent Harness - orchestration layer governing context, tools, tasks, and evaluation.
• Agent Permission Boundaries - safety requirement created by local and external action.
• Agent Skill Supply-Chain Risk - risk introduced when community packages inherit tool and data access.
• Local Agent Execution - deployment pattern that improves control while increasing local responsibility.
• OpenClaw For Teams - enterprise extension requiring identity, memory, audit, and authority separation.
• Delegated Agent Interaction - outcome-oriented interaction pattern OpenClaw makes visible.