The numbers behind the fastest software wave in memory

OpenClaw: from weekend relay to 386,000 stars

OpenClaw did not start as a platform. It was first published on 24 November 2025 as Warelay, a small WhatsApp-to-Claude relay by Austrian developer Peter Steinberger, and was soon renamed Clawdbot (Wikipedia). On 27 January 2026 it became Moltbot after a trademark request from Anthropic over the name's similarity to Claude, the team quipping "Same lobster soul, new shell" (Laravel News, Forbes). Three days later, on 30 January 2026, it was renamed again to OpenClaw (Wikipedia). Three names in ten weeks is the tell: the project was moving faster than its own branding.

• 100,000 GitHub stars in about two months from first publication (TechCrunch, 30 January 2026).
• 247,000 stars by 2 March 2026, about five weeks after its late-January viral breakout (Wikipedia).
• 386,400 stars and 81,200 forks on 16 August 2026 (GitHub, fetched directly for this workshop).

That trajectory makes it one of the fastest-growing open-source projects in history. The stars are a proxy, not a revenue line, but as a measure of developer attention the curve has few precedents.

Hermes Agent: the challenger that took the top rank

Hermes Agent was released on 25 February 2026 by Nous Research under the MIT licence (press reports). It arrived three months after OpenClaw and closed the gap in weeks, not years.

• 231,300 GitHub stars by 16 August 2026 (GitHub).
• Overtook OpenClaw on OpenRouter's global daily rankings as reported on 10 May 2026, at 224 billion daily tokens versus OpenClaw's 186 billion (MarkTechPost).
• Number one daily global rank on OpenRouter as of 16 August 2026, with 36.4 trillion tokens processed in the previous 30 days across 422 models (OpenRouter app page, fetched directly).
• Nous Research reported in talks to raise at least $75 million at a $1.5 billion valuation (TechCrunch, 13 July 2026).

Two projects, neither from a major lab, now sit at the top of a global model-routing marketplace. The lesson for a boardroom is not which agent wins. It is that the category moved from hobby code to infrastructure-scale traffic inside nine months.

The machine traffic crossover, and the price story

The single most important structural datapoint is the crossover. OpenRouter reports that agentic workloads passed human usage on its platform around 1 February 2026, with agentic requests consuming roughly 15 times the tokens per request of human ones (OpenRouter blog). Most traffic through that marketplace is now machines talking to models on behalf of people, not people typing.

Price pressure is part of the same story. DeepSeek roughly doubled its share of OpenRouter weekly token flow, from about 9 percent in January 2026 to about 18 percent by June 2026 (OpenRouter blog). Agents are heavy, repeat consumers of tokens, and heavy consumers shop on price. Cheaper capable models make agents cheaper to run, which creates more agents, which creates more demand for cheap tokens. The flywheel is already turning.

What the analysts say: the sober counterweight

The adoption curves above are real, and so is the friction. The analyst houses are worth quoting precisely because they cut both ways.

• Gartner, on speed: 40 percent of enterprise applications will feature task-specific agents by end 2026, up from under 5 percent in 2025 (Gartner press release).
• Gartner, on failure: separately, over 40 percent of agentic AI projects will be cancelled by end 2027, on cost, unclear business value and inadequate risk controls (Gartner press release).
• McKinsey, State of AI (November 2025): 23 percent of organisations are scaling agentic AI and 39 percent are experimenting, with no single business function above 10 percent scaled.
• IDC: agentic AI is forecast to reach $1.3 trillion, over 26 percent of worldwide IT spending, by 2029.

Sources