The 'Lobster' Creator: Trapped by His Own AI Agent Phenomenon

OpenClaw, the once viral AI agent tool affectionately nicknamed "Lobster," finally rolled out its long-awaited v2026.8.1 update—branded as OpenClaw 2.0—on August 31. This major overhaul, involving 933 contributors (569 first-time submitters) and over 16,000 merged Pull Requests (nearly half the project’s total), aimed to fix critical issues from its predecessor. Key improvements include one-click installation (replacing the complex API Key and command-line setup of 1.0), enhanced memory systems like Active Memory and Background Memory Consolidation (with a "Dream Diary" for off-hours memory sorting), robust security (masked key inputs, trust boundary checks), and shared cloud sessions for collaboration.

Yet, OpenClaw’s popularity has plummeted. GitHub visits dropped from 28.8 million in April to under 10 million pre-update, while weekly new stars fell from 5,900/day in February to 800/week—far behind Hermes Agent’s 9,000 weekly stars. NPM downloads, once 1.5 million in mid-March, are back to pre-boom levels. Steinberg’s new project Crabbox has also flopped, with total stars less than OpenClaw’s daily growth at its peak.

In an August YC podcast, Steinberg revealed his struggles: his private phone leaked, he got late-night journalist calls, and faced constant "OpenClaw is dead" articles. Joining OpenAI made OpenClaw feel "no longer fun"—he shifted from adding features to fixing bugs, handling security incidents, and managing AI-generated Pull Requests that distracted him from core work.

Steinberg admitted a strategic mistake: optimizing OpenClaw deeply for Anthropic’s Claude. In February, Anthropic banned using Claude subscriptions via third-party tools like OpenClaw; in April, it cut access entirely, forcing users to pay higher API costs. This hit OpenClaw hard, as most users relied on Claude.

Competitors capitalized: Nous Research’s Hermes Agent (launched Feb) fixed OpenClaw’s pain points (one-click install, self-evolving skills) and gained traction. Others like KimiClaw (cloud-hosted), CoPaw (Alibaba’s dual deployment), QClaw (Tencent), and Moltbook (ByteDance) entered the fray.

Ironically, Steinberg’s methods influenced rivals like Codex: it adopted OpenClaw’s multi-agent parallelism and long task loops, added computer use and cross-session memory, and implemented Codex Security (pre-auditing task risks) from OpenClaw’s mistakes.

Sources

• Guancha.cn (User-generated content, ID: 1726955)