OpenClaw

OpenClaw is an open agent platform created by Peter Steinberger that runs on user-chosen infrastructure and works through existing messaging apps. Its assistant combines chat access with the ability to write and run code. On July 8, 2026, Dave Morin and Steinberger announced the OpenClaw Foundation as an American 501(c)(3) nonprofit stewarding the software as MIT licensed, open, and independent, with its first full-time team.[8][11][10]

That foundation followed a project Steinberger began in late 2025, before the OpenClaw name and its paid maintenance structure.[8][10]

The project began as “WhatsApp Relay,” a weekend experiment. Clawd emerged in November 2025 as a pun on Claude. According to Steinberger, Anthropic’s legal team asked him to reconsider the name; Moltbot followed, chosen with the community during a 5am Discord brainstorm. On January 29, 2026, he announced OpenClaw after trademark searches, domain purchases, and preparation of migration code.[8]

OpenClaw’s January product description emphasized deployment on a laptop, homelab, or VPS and access through WhatsApp, Telegram, Discord, Slack, and Teams. The rebranding release added Twitch and Google Chat plugins, KIMI K2.5 and Xiaomi MiMo-V2-Flash support, and image sending in web chat. Steinberger reported that the repository had over 100,000 GitHub stars and that the project drew 2 million visitors in a single week; these were repository and traffic figures, rather than active-user counts. He was adding maintainers, establishing processes for incoming pull requests and issues, and exploring compensation, preferably for full-time work.[8]

In January, OpenClaw was built on components from Pi, the coding agent written by Mario Zechner. Armin Ronacher, who identified Steinberger as a friend, explained the relationship in Pi: The Minimal Agent Within OpenClaw. Pi had four core tools—Read, Write, Edit, and Bash—and an extension system that could persist state into sessions. OpenClaw connected this code-writing and code-running approach to a messaging channel.[11]

Pi’s extension architecture supported a write–reload–test loop: an agent could change an extension, hot-reload it, and keep testing until it worked. Its tree-shaped sessions allowed branching to repair a tool, then returning to an earlier context with a summary of the branch. Sessions could contain messages from different model providers, although Ronacher noted limits to portability between providers. Pi lacked built-in MCP support; OpenClaw used mcporter to expose MCP calls through a command-line interface or TypeScript bindings.[11]

On February 14, 2026, Steinberger announced that he was joining OpenAI and intended to move OpenClaw into a foundation while keeping it open and independent. He described his next mission as building “an agent that even my mum can use.” Having already spent 13 years building a company, he said he preferred broader access to agents over creating another large company. He reported that OpenAI already sponsored OpenClaw and had committed to enabling his continued project work.[9]

The July foundation announcement gave that plan an organizational structure for governance, stable funding, and paying maintainers. Steinberger retained decision-making authority, especially on technical matters. The initial full-time engineering team comprised Vincent Koc as Chief Architect, Josh Avant, Patrick Erichsen, Dallin Romney, Jason Sy, and Gideon Adegbesan. Operations staff were Jen Vescio for partnerships, Matt Jasie for finance, Hannes Rudolph for community, and Kelly Pike for talent.[10]

The foundation named Offline Holdings, Lobster Computer Company, University of Michigan, and OpenAI as major donors, describing Michigan as its largest donor. It also reported OpenAI support for inference and platform hardening through Codex Security. Claw Labs was a separate team inside OpenAI, led by Steinberger and working on shared product improvements. The announced support took the form of sponsorship and donations, with no amounts specified.[9][10]

ClawHub, OpenClaw’s skill marketplace, extended the agent’s capabilities while introducing code and instructions with access to its tools and data. On February 7, Steinberger, Jamieson O’Reilly, and Bernardo Quintero announced a VirusTotal partnership to scan published skills. OpenClaw also brought on O’Reilly as lead security advisor.[15]

The scanning pipeline packaged skills into deterministic ZIP bundles, computed a SHA-256 hash, and checked VirusTotal for an existing Code Insight verdict. Bundles without a usable analysis were uploaded for fresh scanning. Gemini-powered Code Insight examined SKILL.md and referenced scripts and resources. Benign verdicts triggered approval; suspicious verdicts added warnings while leaving skills available; malicious verdicts blocked downloads. Active skills were re-scanned daily. The partners explicitly cautioned that a clean scan did not establish safety, particularly against carefully crafted prompt injection.[15]

ClawHub Security Signals, submitted on May 31, 2026, studied disagreement among VirusTotal, static heuristic analysis, and NVIDIA SkillSpector across 67,453 latest public OpenClaw skill versions. Its authors were Vincent Koc, Patrick Erichsen, Jacob Tomlinson, Agustin Rivera, Michael Appel, and Nir Paz. They reported that 81.9% of flagged skills were identified by only one scanner, while 0.69% of corpus skills were flagged by all three. The study examined differing security signals, not malicious-skill prevalence: its labels were automated registry verdicts rather than human-annotated ground truth.[17]

NVIDIA announced NemoClaw for OpenClaw on March 16, 2026. The NVIDIA stack was designed to install Nemotron models and the OpenShell runtime in one command. NVIDIA described an isolated sandbox with policy-based security, network, and privacy controls, plus a privacy router for accessing cloud frontier models alongside locally running open models. These were announced design capabilities; NVIDIA qualified the availability and timing of products and features in the release.[16]

By July, the foundation reported collaboration with Microsoft on Scout agents powered by OpenClaw technology and on Windows integration. Red Hat contributed enterprise open-source expertise, while Tencent supplied full-time maintainers for security, stability, and ClawHub. Foundation-convened councils addressed agent identity, agent profiles, evaluations, and enterprise deployment.[10]

The foundation’s July announcement reported day-to-day work with more than 30 organizations. It also reported 34 ClawCon events across 16 countries over five months, with nearly 30,000 signups. Those figures describe collaboration and event registration, rather than customers or attendance.[10]