OpenClaw: What It Is, Why Anthropic Ended Claude-Sub Coverage, and the 369k-Star Ecosystem

OpenClaw is the agent project most teams have heard about and not actually understood. 369,000+ GitHub stars, a top-tier Lex Fridman episode, a name change forced by Anthropic's lawyers, a founder who jumped to OpenAI three days after the podcast aired, and an April 2026 subscription ban that turned the project into a commercial flashpoint — all in roughly 180 days.

This page is the honest summary. What OpenClaw is, what it isn't, what its ecosystem looks like, and where it fits next to Claude Code, Cursor, Codex, and the other tools RuleSell's multi-tool catalog serves.

What OpenClaw actually is

OpenClaw is an open-source always-on AI agent that runs on a user's local machine and processes a multi-channel inbox — chat messages, emails, files dropped into watched folders, and scheduled "heartbeat" tasks. The agent reads its identity from a small set of workspace files (SOUL.md, AGENTS.md, TOOLS.md, MEMORY.md, HEARTBEAT.md) and uses whatever model backend the user wires up: Claude, GPT, Gemini, or local open models.

The project's tagline — "the lobster way" — refers to its founder Peter Steinberger's framing of agents as digital crustaceans that scuttle through your filesystem doing assigned work. The Wikipedia entry describes it as a "digital employee" that runs "on a user's local machine" (en.wikipedia.org/wiki/OpenClaw). That framing is more useful than "AI agent" because it captures the long-running, file-rooted, schedule-driven shape — which is genuinely different from request-response coding assistants like Claude Code or Cursor.

OpenClaw launched in November 2025 as Clawdbot, was renamed to Moltbot on January 27 2026 after an Anthropic trademark complaint about "Clawd" being phonetically too close to "Claude," and reached its final name OpenClaw three days later on January 30 2026 (Lexology coverage). The double rename in 72 hours is part of the project's lore.

By March 2026, OpenClaw had 247,000 stars and 47,700 forks. By May 2026, the star count had crossed 369,000+ — making it the most-starred new AI project of the 2025-2026 cycle. The Lex Fridman Podcast episode #491, published February 11 2026 and titled "OpenClaw: The Viral AI Agent that Broke the Internet," was the discovery event for most non-Twitter developers.

Three days after the podcast, on February 14 2026, Steinberger joined OpenAI and a foundation was set up to steward the project. That foundation is the entity Anthropic later moved against.

Why Anthropic banned OpenClaw from Claude subscriptions

In April 2026, Anthropic blocked OpenClaw from authenticating against Claude Pro and Claude Max subscriptions, pushing OpenClaw users onto per-token API billing instead. Anthropic has not published a full rationale; the reporting frames it as terms-of-service enforcement against using consumer-tier subscriptions to power headless, always-on agent workloads.

The legal reading is straightforward — TOS for Pro/Max subscriptions does not anticipate 24/7 agent traffic. The commercial reading is more interesting. Anthropic shipped Managed Agents the same month (thenewstack.io), priced at "$0.08 per session-hour" — direct competition with the kind of always-on workload OpenClaw represents. The takeaway: agent runtimes are now treated as a competitive surface, not just a usage pattern.

This matters for anyone building on top of a single vendor's runtime. A tool that authenticates against a subscription you don't control can lose its login overnight. That's part of why RuleSell's catalog is multi-tool by design — rulesets published here are written to work across Claude Code, Cursor, Codex CLI, Cline, Aider, Gemini CLI, Warp, Windsurf, and the rest. When one runtime changes its terms, your asset still has nine homes.

The ecosystem: NemoClaw, NanoClaw, ZeroClaw, ClawHub

OpenClaw spawned a fast-moving ecosystem of forks, distributions, and adjacent tools. The main pieces, sourced from the InnFactory ecosystem map:

• NemoClaw — NVIDIA's distribution, optimized for local GPU inference and zero-trust credential isolation (VentureBeat coverage)
• NanoClaw — a stripped-down build for embedded and edge use
• ZeroClaw — a zero-dependency variant for security-conscious environments
• Moltis, PicoClaw, IronClaw, MicroClaw — smaller forks targeting specific niches
• ClawHub — the community hub where OpenClaw users share skills and tools

This isn't a knock — community hubs are valuable and serve real needs. But it's a different category from what RuleSell does. RuleSell is the multi-tool, paid marketplace with Stripe Connect creator payouts, affiliate revenue share, and a founding-creator program. Different model, different audience, different value to creators. The two coexist; they aren't direct substitutes.

ClawHub's other relevant data point: Snyk's ToxicSkills report (snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub), published February 5 2026, audited 3,984 ClawHub skills and found 13.4% with critical issues plus 76 active malicious payloads — under the name ClawHavoc. The trust problem is structural for any free-publish hub. It's also the reason RuleSell invested in trust infrastructure up front.

Where this fits next to Claude Code, Cursor, Codex

OpenClaw and the coding-focused tools — Claude Code, Cursor, Codex CLI, Cline, Aider, Warp 2.0 — solve different shapes of problem. The honest comparison:

• Coding agents (Claude Code, Cursor, Codex, Cline, Aider, Warp): you have a repo, you want changes. Request-response shape. Edits, tests, commits.
• OpenClaw: you have an inbox and a calendar and a watched folder, you want it tended. Long-running shape. Triage, response, scheduled work.

The portable asset across all of these is the Markdown config file. AGENTS.md, CLAUDE.md, SOUL.md, .cursorrules, .windsurfrules — the format varies, the substance overlaps heavily. A well-written agent-rules file is reusable across most of them with minor adaptation. That portability is what makes a multi-tool catalog work in the first place.

Where this fails / what we don't know

A few things this page can't tell you, that future work should:

• No direct ClawHub audit. The counts above (52k tools, 12M downloads, 3,984 audited skills) come from secondary reporting. We have not fetched ClawHub directly, sampled its category mix, or counted active creators. The numbers are reasonable but not first-hand. A 30-minute audit would tighten this.
• Anthropic's full rationale for the subscription block. Public coverage paraphrases the policy reason. Anthropic has not posted a detailed engineering or policy rationale, so the reading above is inferred from the timing (Managed Agents the same month) plus public TOS.
• Star count vs. active use. 369k stars is impressive but doesn't measure how many developers run OpenClaw daily versus how many starred it after the Lex Fridman episode. We have no daily-active-user number. Treat the star count as awareness, not adoption.
• Long-term legal exposure on the rename. The Clawdbot → Moltbot → OpenClaw chain settled the immediate trademark dispute, but the broader question of whether agent-runtime names that pun on model names will survive future legal scrutiny is open.

What to read next

• /topic/cli-everything-wave — the broader CLI-everything meta-trend OpenClaw is part of
• /topic/agent-skills-cross-tool — how SKILL.md became a cross-tool standard
• /topic/dual-wielding-codex-claude-code — the multi-CLI stack pattern
• /topic/agents-md — the cross-tool rules file standard
• /for/anthropic-managed-agents — Anthropic's own managed agent runtime
• /for/claude-code — the coding-agent stack OpenClaw complements

Sources

• OpenClaw repository — github.com/openclaw/openclaw. 369k stars as of May 2026.
• Wikipedia, "OpenClaw" — en.wikipedia.org/wiki/OpenClaw. Creator, dates, rename chronology.
• Lex Fridman Podcast #491, "Peter Steinberger" — lexfridman.com/peter-steinberger. Published Feb 11 2026.
• TechCrunch, "OpenClaw creator Peter Steinberger joins OpenAI" — techcrunch.com. Feb 15 2026.
• Lexology, OpenClaw trademark coverage — lexology.com. Rename details.
• Revolution in AI, Anthropic subscription ban coverage — revolutioninai.com. April 2026.
• The New Stack, "Anthropic Managed Agents: Dreaming Outcomes" — thenewstack.io. Same-month context for the ban.
• InnFactory, OpenClaw ecosystem map — innfactory.ai. NemoClaw, NanoClaw, ZeroClaw, ClawHub, tool counts.
• Snyk, "ToxicSkills: Malicious AI Agent Skills on ClawHub" — snyk.io. Feb 5 2026.
• Capodieci on Medium, "OpenClaw workspace files explained" — capodieci.medium.com. SOUL.md, AGENTS.md, MEMORY.md, HEARTBEAT.md.
• VentureBeat, "AI agent zero-trust architecture" — venturebeat.com. NemoClaw context.