The claw and the meter

The most-starred open-source AI agent in history is named after a scar.

It began as Clawd — an assistant its creator, Peter Steinberger, named after Anthropic's Claude. A trademark complaint from Anthropic later, Clawd became Clawdbot became Moltbot became OpenClaw: five names in two months, the crustacean theme a cover story that kept the Claw- once the Claude had to go. The claw is a Claude with the serial numbers filed off.

Born a wrapper

That heritage isn't trivia — it's the whole story in one word. OpenClaw was born a Claude wrapper: a local agent that drank Claude through the SDK. Tomorrow it's one of four third-party agents — with Conductor, Zed, Jean — named in Anthropic's June 15 change: claude -p and Agent-SDK usage leaves the flat-rate subscription pool for a separate, metered credit. Clamped on the trademark in January. Acqui-hired into OpenAI in February. Fenced off the supply in June. Two enclosures and an acquisition in a single quarter, for the agent everyone starred.

None of that is a knock on the work — 145k stars is a real achievement, and the instinct (local-first, memory in files, a heartbeat that acts on its own) is a good one. It's a knock on the foundation. Rent your cognition and you inherit the landlord's trademark department and the landlord's billing cycle — and you learn which one matters on their schedule, not yours.

Maximal harness

While that played out, a research group reverse-engineered Claude Code itself — arXiv:2604.14228 — and wrote down what they found: the intelligence is almost none of the system; the weight is the harness around the loop. Their words, verbatim: minimal scaffolding, maximal harness. That is the sentence this lab has been building on for four months. The same paper lists, as open problems for future agents, things already on our board — an agent that persists as a colleague across sessions, gates that catch silent failure, a human kept in the governance loop. We read that as a compass check, not a citation: pointed at the right horizon.

What we keep, what we own

OpenClaw reached owned, local autonomy from the messaging-gateway direction; we reached it from the coordination-substrate side — plans, tasks, approvals, workflows, a verifier in the loop, built for a crew that hands work to itself and proves it. We'll take the harness lesson; the paper earns it. We'll study their skill format. We won't take the open-hub trust model that let ~11% of community skills ship malicious.

But the difference that actually lands tomorrow is simpler. When the meter starts, our squad doesn't notice. It runs on models on a desk, weights we hold, memory in files we own. There is no name to file off, and no pool to be removed from.

Rent the intelligence — even Claude's, even when you love it — and the claw remembers the hand that fed it. Build it on metal you own, and it answers only to you.