Preface

In 2026, one name was almost everywhere in the open source community: OpenClaw. This personal AI assistant project initiated by Austrian developer Peter Steinberger has accumulated more than 380,000 stars on GitHub, and is widely recognized by official and community as one of the fastest-growing open source repositories in recent years. Unlike cloud-based assistants such as ChatGPT and Claude, OpenClaw emphasizes local-first, self-hosted, and data stays on your own devices – you can run it on a Mac mini, Linux server, or even a Raspberry Pi, then chat with it via daily messaging tools like WhatsApp, Telegram, and Discord, letting it execute Shell commands locally, control browsers, read and write files, and run scheduled tasks.

In July 2026, the project took another important step: launching the extended-stable update channel and launching the public Maturity Scorecard, paving the way for formal LTS long-term support. Based on public information from the OpenClaw official repository, official website, and blog, this article sorts out why this project has attracted so much attention, and how developers can get started with it.

What It Is

OpenClaw is an open source personal AI assistant framework with the slogan “Your own personal AI assistant. Any OS. Any Platform.”. The project is hosted at github.com/openclaw/openclaw, and the official website is openclaw.ai.

From an architectural perspective, OpenClaw is not a large language model itself, but an Agent Harness (intelligent agent runtime): it connects the LLM of your choice (OpenAI, Anthropic, local models, etc.), and then gives the agent a set of “tools” – terminal execution, browser automation, file system access, scheduled tasks (cron), persistent memory, and various messaging channels accessed via the Gateway. Context and Skill configurations are saved locally by default, not in a closed account on a certain SaaS platform.

The project uses the MIT license, and anyone can use, modify, and redevelop it for free. The repository was created in November 2025, and the star count exceeded 300,000 in just a few months. As of early August 2026, it has nearly 390,000 stars, and the number of forks has exceeded 80,000, with extremely high activity in Issues and PRs.

Why It Became a Hit in 2026

OpenClaw’s popularity essentially comes from hitting several developer pain points while delivering solutions:

  1. Local Control: Model API keys, conversation memories, and Skill files are all on your own machine, which is suitable for individuals and small teams who are sensitive to data sovereignty.
  2. “Chat and Give Commands”: No need to open an IDE or Web UI specifically; send a message in Telegram, and the agent can run tests, check emails, and generate PRs in the background.
  3. Extensibility: Through the Skill mechanism and the ClawHub community repository, capabilities can be installed, versioned, and shared just like npm packages.
  4. Ecosystem Endorsement: Public interactions between large factories such as Microsoft Build, OpenAI, NVIDIA, and the founder further amplified the project’s voice; channels such as Y Combinator also called it “one of the software repositories with the fastest star growth on GitHub”.

The keywords that repeatedly appear in community feedback are: persistent memory, active heartbeat, cross-channel context, and “the agent can really help you work instead of just chatting”. These experiences combined have quickly turned OpenClaw from a weekend hack project into a daily infrastructure for many developers.

Founder and OpenClaw Foundation

The creator of OpenClaw is Peter Steinberger (GitHub / X: @steipete). He was previously known to mobile developers for founding the PDF framework company PSPDFKit (now Nutrient). At the end of 2025, Steinberger released the predecessor of OpenClaw as a personal experimental project (the community once briefly used names such as Clawdbot and Moltbot, which were later unified as OpenClaw).

In February 2026, Steinberger announced that he would join OpenAI, while emphasizing that OpenClaw would continue to evolve independently in open source form. On July 8, 2026, he and Dave Morin and others jointly announced the establishment of the OpenClaw Foundation – a 501(c)(3) non-profit organization whose mission is to govern the project for the long term under the MIT license and avoid being “acquired and turned closed-source” by a single commercial entity. After the foundation was established, Steinberger is still responsible for the technical direction; OpenAI, NVIDIA, Microsoft, Tencent, etc. are listed as partners or donors, but the official clearly stated that the ownership and open source commitment of OpenClaw are guaranteed by the foundation.

For ordinary users, this means: even if the founder’s personal career path changes, the project itself will have an institutional governance structure, similar to the Mozilla Foundation’s positioning for Firefox.

Multi-channel Message Routing

One of the highlights of OpenClaw is the Gateway + Channels architecture: the Gateway acts as a hub, standardizes messages from different IM platforms and hands them to the Agent for processing, then routes the replies back to the original channel. The channels listed in the official documentation include WhatsApp, Telegram, Discord, Slack, Signal, iMessage, WebChat, IRC, etc.; Microsoft Teams, Matrix, Google Chat, Feishu, LINE, etc. can be installed as needed in the form of official plugins or ClawHub plugins. The community documentation (v2026.6) counts approximately 26 documented channels, and continues to expand through plugins.

WhatsApp is one of the most frequently mentioned integrations and one of the most installed channels. It is implemented through the Baileys library, requires scanning a code to pair, and supports private chats and groups, media messages, and read receipts. The typical configuration process is as follows:

# Install the official WhatsApp plugin
openclaw plugins install @openclaw/whatsapp

# Scan the code to log in
openclaw channels login --channel whatsapp

In openclaw.config.json, you can set DM policies for WhatsApp (pairing / allowlist / open), group whitelists, media size limits, etc. Multiple channels can be online at the same time, and OpenClaw automatically routes them according to the session source, without the need to deploy separate Agent instances for each platform.

Telegram is often recommended as the first experience channel because it only requires a Bot Token and is the easiest to get started with; WhatsApp configuration is slightly more complex, but it is close to the daily habits of domestic users, so the discussion volume is also very high.

What Self-hosted Agents Can Do

After running OpenClaw locally or on your own VPS, the agent has roughly the following capabilities:

  • Shell / Code Execution: Run commands in a sandbox or local terminal, and cooperate with toolchains such as Claude Code and Codex CLI for automated development.
  • Browser Automation: Open web pages, fill out forms, and crawl information through Skills such as Playwright.
  • Files and Memory: Read and write workspace files, and retain personas and user preferences across sessions.
  • Scheduled and Background Tasks: Run inspections, daily reports, and monitoring alerts triggered by cron expressions.
  • Webhook Integration: Receive events from Sentry, GitHub, etc., and automatically create Issues or open PRs.

For installation, the official provides a one-click script that supports macOS, Linux, and Windows (on Windows, you can also use Microsoft Execution Containers for sandbox isolation):

curl -fsSL https://openclaw.ai/install.sh | bash

The script will install Node.js and the OpenClaw CLI. Then run openclaw onboard to complete the model configuration, channel binding, and Skill selection interactively. If you have already installed npm, you can also directly use:

npm install -g openclaw@latest

Skill Ecosystem and ClawHub

The extension unit of OpenClaw is Skill: each Skill is a directory, and the core file is SKILL.md (including YAML metadata and operation instructions for the Agent), which is compatible with the AgentSkills specification. Skills can declare dependent environment variables, binary tools, and OS restrictions; OpenClaw will filter according to the current environment when loading, preventing the Agent from calling unavailable capabilities.

There are three layers of Skill sources, sorted from lowest to highest priority:

  1. Bundled Skills: Basic capabilities built into the npm package or OpenClaw.app.
  2. Managed Skills: Local overrides under ~/.openclaw/skills.
  3. Workspace Skills: The /skills directory of the current workspace, with the highest priority.

ClawHub (clawhub.ai, GitHub repository openclaw/clawhub) is the public Skill registry for OpenClaw, similar to npm. Developers can search, install, publish, and version Skills; the registry supports semantic search based on embedding, not just keyword matching. Common commands:

# Install the Skill to the current workspace
clawhub install <skill-slug>

# Update installed Skills
clawhub update --all

# Publish your own Skill (requires clawhub login first)
clawhub publish

Community reports show that as of early 2026, ClawHub has收录 more than 13,000 Skills, covering scenarios such as GitHub operations, browser automation, smart home, and SEO writing. This is also the core of OpenClaw’s difference from “only chat” assistants: the capability boundary is continuously expanded by the community.

Moving Towards LTS: extended-stable and the Maturity Scorecard

The side effect of rapid iteration is that early users often complained “it crashes every time you update”. Around July 30, 2026, the OpenClaw official blog published the article On the Road to LTS: Extended-Stable Releases and the Maturity Scorecard, announcing two mechanisms for production users.

The extended-stable channel is not a full LTS, but a long-term stable branch released monthly: the first version is OpenClaw 2026.6.33 (based on 2026.6.11, with backported security and reliability fixes from subsequent versions). The naming rule starts with YYYY.M.33, and the patch number increases with backports; each extended-stable version will be supported at least until the next monthly version is released.

Installation and switching methods:

npm install -g openclaw@extended-stable

# Permanently update the channel
openclaw update --channel extended-stable

The Maturity Scorecard (public page: docs2.openclaw.ai/maturity/scorecard) inventory the functional aspects of OpenClaw: the official stated that it covers 50 Surface (product capability areas), 281 Capability Area (capability domains), and gives grades such as Stable / Beta / Experimental based on quality, completeness, and QA evidence. Mature functions will receive priority Issue handling and end-to-end test coverage; the team’s goal is to make the E2E test coverage of Stable-level functions exceed 90% (this is a goal, not a full guarantee at present).

For teams deploying in production or semi-production environments, the official recommends: run critical workflows on extended-stable, and keep experimental functions on the latest/dev channel; check the scorecard before going online to confirm that the Surface used has reached Stable level.

Quick Start Tips

If you want to experience OpenClaw locally, you can follow the following steps:

  1. Prepare a machine that is always on (old Mac, Linux VPS, or Raspberry Pi are all acceptable), and ensure that Node.js 18+ is available.
  2. Run the official installation script or npm install -g openclaw@latest.
  3. Run openclaw onboard to configure the LLM API Key (you can also bind a ChatGPT subscription account, and the official already supports OAuth login).
  4. Prioritize accessing Telegram for a smoke test; after getting familiar with it, install the WhatsApp plugin and scan the code to pair.
  5. Search for required Skills on ClawHub (such as GitHub, browser, email) and install them with clawhub install.
  6. If you plan to run it for a long time, switch to the extended-stable channel, and pay attention to the level of the corresponding functions in the maturity scorecard.

The learning curve of OpenClaw is not smooth: CLI configuration, Gateway processes, plugins, and Skill permissions all require a certain degree of operational awareness. But once you get it running, you will get a personal agent that runs 24/7 on your own turf – this is also why it has continued to occupy the center of developer topics in 2026.

References