OpenClaw is an open-source, self-hosted AI assistant that runs on your own computer and is used through chat apps such as WhatsApp, Telegram, Slack and iMessage; it was previously called Clawdbot and then Moltbot before taking its current name on 30 January 2026. This profile collects the published facts in one place: who stewards it, how it is licensed and priced, its naming history, release channels, which model providers it connects to, which channels and platforms it supports, where it stores data and credentials, and the security settings its documentation describes. It is written for individuals deciding whether to run it, and for IT, security and governance teams who need to understand it before it appears on managed devices. Every fact comes from the OpenClaw website, its GitHub repository and its documentation, listed under Standards and references. Where those sources say nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses OpenClaw's public documentation as of November 2026, and is not sponsored or reviewed by the OpenClaw Foundation. Product features change, so confirm settings against OpenClaw's current documentation.
Key takeaways
- Open source, MIT licensed, no paid tier. The OpenClaw Foundation, an independent US 501(c)(3) non-profit, stewards the project; the website states "No subscription. No hosted tier. No token."
- Three names in ten weeks. The project's own lore page traces the assistant as Clawd in Clawdbot from 25 November 2025, Moltbot from 27 January 2026 after a trademark request from Anthropic, and OpenClaw from 30 January 2026.
- Bring your own model. OpenClaw is not a model: it connects to hosted, subscription-backed, gateway or local models, with official provider plugins for OpenAI, Anthropic, Google Gemini, Z.AI and more than 25 other providers.
- A local Gateway is the control plane. One Gateway process manages sessions, tools and channel connections; companion apps cover macOS, Windows, Linux, iOS and Android.
- Customer-side settings decide the risk. The Gateway binds to loopback and unknown DM senders get a pairing code by default; tool sandboxing is off by default and is turned on in configuration.
What is OpenClaw?
OpenClaw describes itself as "the open-source AI assistant that runs on your machine and works from WhatsApp, Telegram, or any chat app." The repository README says it "runs on your own computer and meets you in the channels you already use", with native apps for macOS, iOS, Android, Windows and Linux. It belongs to the class of software usually called personal AI agents: it keeps persistent memory, browses and fills in forms, reads and writes files, runs shell commands, follows scheduled automations and extends itself through skills and plugins. For how this class of product compares across vendors, see Personal AI Agents.
The README states that "State, memory, and credentials live on your hardware" and that models and agent harnesses "are plugins you can swap without changing anything else." The project was created by Peter Steinberger and its community, according to the README. The website states that OpenAI is a donor and that Steinberger works there, and that "OpenClaw is not an OpenAI product." The Foundation employs the core team and signs releases; donors listed in the README include Amazon, Lobster Computer Company, Offline Holdings, OpenAI, Red Hat and the University of Michigan, with infrastructure support from Blacksmith, Convex, GitHub, NVIDIA and Vercel. The README says none of them "own or direct the project".
Key facts at a glance
As of 2026-11-15.
| Field | Value | Source |
|---|---|---|
| Product type | Self-hosted personal AI assistant (agent), used through chat channels and companion apps | openclaw.ai |
| Steward | OpenClaw Foundation, independent US 501(c)(3) non-profit | openclaw.ai |
| Former names | Warelay (WhatsApp gateway), Clawdbot, Moltbot | OpenClaw lore |
| Licence | MIT, copyright OpenClaw Foundation | GitHub README |
| Price | Free; no subscription, hosted tier, enterprise edition or paid version | openclaw.ai |
| Core language and runtime | TypeScript; Node.js is the default runtime, Bun 1.4.4+ as an opt-in | Platforms |
| Node.js versions | 24.16.0+ on 24.x or 26.1.0+; 24.22.1+ or 26.12.1+ "strongly recommended for security" | GitHub README |
| Latest desktop releases listed | macOS v2026.11.3 (macOS 15+), Windows v2026.11.1 (Windows 10/11, x64 or ARM64), Linux v2026.11.2 (x64, AppImage or Debian package) | openclaw.ai |
| Versioning | Calendar style, for example 2026.11.3; 140 release entries in the changelog index, from 2026.1.5 | Changelog |
| Models | Bring your own: hosted, subscription-backed, gateway or local | Model providers |
| Skills registry | ClawHub (clawhub.ai) | GitHub README |
| Telemetry | Daily update check by default; anonymous feature statistics off by default | Telemetry |
Release and version history
The project's naming history is recorded on its own lore page in the documentation, written in a light-hearted voice; the dates below are the ones it states.
| Date | Event | Source |
|---|---|---|
| Before November 2025 | The project begins as Warelay, described as "a sensible name for a WhatsApp gateway" | OpenClaw lore |
| 25 November 2025 | The assistant is known as Clawd, living in Clawdbot (Clawd period stated as 25 November 2025 to 27 January 2026) | OpenClaw lore |
| 27 January 2026 | After Anthropic asked for a name change over trademark concerns, the community picks Moltbot | OpenClaw lore |
| 30 January 2026 | Renamed OpenClaw; the GitHub repository moves to github.com/openclaw/openclaw, npm packages are released under the new name and docs move to docs.openclaw.ai | OpenClaw lore |
| 2026 | Releases follow a calendar scheme; the changelog index runs from 2026.1.5 to 2026.11.3 | Changelog |
| 2026.8.1 | Release adding conversation search and sessions that run on paired devices or cloud workers | 2026.8.1 notes |
| 2026.11.3 | Security patch release with no new features; fixes certificate verification for worker transfers with blank fingerprints and reduces credential exposure | 2026.11.3 notes |
Release channels
- stable: npm dist-tag latest, recommended for most users.
- extended-stable: a trailing, supported-month package channel that never applies automatically.
- beta: the newest version across the beta and latest dist-tags; stable builds usually ship here first.
- dev: the moving head of main, which the documentation says not to run for production gateways.
Source: Release channels.
How much does OpenClaw cost?
| Item | Price | Notes | Source |
|---|---|---|---|
| OpenClaw software | Free | MIT licence; "No subscription. No hosted tier. No token." | openclaw.ai |
| Enterprise or paid edition | None | The Foundation text states "There is no enterprise edition and no paid version" | openclaw.ai |
| Model usage | Set by the model provider you connect | Billed by that provider under its API or subscription terms; local models need your own hardware | Official provider plugins |
| Hosting | Your own device or a server you choose | The docs list VPS guides for Azure, Daytona, EasyRunner, exe.dev, Fly.io, GCP and Hetzner | Platforms |
An assistant that runs scheduled tasks all day can use many tokens. Current model rates are in the LLM Pricing Tracker; set a spend limit on every key OpenClaw uses.
Supported models and providers
OpenClaw supports models through provider plugins. Official provider plugins publish their own model catalogue, so the operator enables the plugin, sets authentication and picks a model; other providers are added through models.providers with a custom base URL. The documentation names the following, as of 2026-11-15.
| Provider route | Details from the docs | Source |
|---|---|---|
| Anthropic | Provider anthropic, ANTHROPIC_API_KEY, example model anthropic/claude-opus-5-5; Claude CLI reuse is described as a sanctioned integration path | Official provider plugins |
| Gemini by API key, plus Google Vertex and a Gemini CLI runtime | Official provider plugins | |
| OpenAI | Provider openai, OPENAI_API_KEY; fresh setups default to openai/gpt-6-astra; ChatGPT and Codex OAuth sign-in also supported | Official provider plugins |
| Z.AI (GLM), Vercel AI Gateway, OpenCode | Each has its own official provider plugin section | Official provider plugins |
| Other bundled provider plugins | 30 listed, including BytePlus, Cerebras, Cohere, DeepInfra, DeepSeek, GitHub Copilot, Groq, Hugging Face Inference, MiniMax, Mistral, Moonshot, NVIDIA, OpenRouter, Qianfan, Tencent TokenHub, Together, Volcano Engine (Doubao), xAI and Xiaomi | Official provider plugins |
| Custom providers and local runtimes | Moonshot AI (Kimi) and Kimi Coding, Synthetic, MiniMax, llama.cpp, llmman, LM Studio, Ollama, vLLM, SGLang, and local proxies such as LiteLLM | Model providers |
Provider plugins accept several keys for rotation (for example OPENAI_API_KEYS or numbered variables), and a model failover feature handles fallback chains, retries and auth-profile cooldowns. Adding a provider's credentials does not change the primary model; that is changed deliberately with the models set command or an explicit config entry. The prompt injection page notes that resistance "is not uniform across model tiers" and that smaller or cheaper models are more susceptible to tool misuse, so model choice is also a security setting. For the models themselves, see Frontier AI Models Compared.
Channels, platforms and tools
- Chat channels: the website refers to "any of its 29 channels"; the documentation's generated channel catalogue lists 33 entries, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Google Chat, Microsoft Teams, Matrix, Mattermost, Feishu, LINE, QQ bot, WeChat and WeCom (external plugins), SMS through Twilio, X, WebChat and an A2A channel for external agents. Telegram is the suggested first channel because it needs only a bot token.
- Gateway: the local control plane for sessions, tools, events and channel connections, managed from the Control UI, CLI or terminal UI. It installs as a LaunchAgent on macOS, a systemd user service on Linux and WSL2, and a Scheduled Task on Windows.
- Companion apps and nodes: Linux, Windows Hub (recommended on Windows, with a Tauri app as backup), a macOS menu bar app and iOS and Android nodes, adding voice, Canvas, camera, screen and device-local actions.
- Skills: markdown skill files load from workspace, project, personal, managed and bundled locations in a set precedence order, with workspace skills highest; community skills are shared through ClawHub. See the Agent Skills Security Guide.
- Integrations named on the website: browser, Gmail, GitHub, Obsidian, X, Spotify and Hue.
Where is OpenClaw available?
| Channel | Details | Source |
|---|---|---|
| Installer scripts | macOS, Linux and Windows; provisions a supported Node.js runtime when needed and starts onboarding | GitHub README |
| npm package | openclaw on npm for users who manage Node.js themselves; pnpm and Bun also supported | GitHub README |
| Desktop apps | macOS, Windows and Linux builds listed on the website | openclaw.ai |
| Mobile | iOS and Android companion nodes | Platforms |
| Other installs | Docker, Nix and source builds (pnpm workspace); ChromeOS (Crostini) and Omarchy guides | GitHub README, Platforms |
Access restrictions
There is no account, waitlist or regional restriction on the software stated on the pages checked; anyone can install it under the MIT licence. Access to models depends on the provider the operator connects, and access to each chat channel depends on that platform's own bot or account rules. Within a deployment, access is controlled by OpenClaw's own settings: DM pairing, allowlists, mention gating in groups and per-agent tool policies, described below.
Security model and defaults
The security documentation says OpenClaw "ships with conservative defaults". These settings are the customer's responsibility to keep or tighten.
| Area | Documented default or setting | Source |
|---|---|---|
| Network binding | On a regular host install the Gateway binds to loopback; the hardened baseline also sets token authentication (gateway.auth.mode "token") | Security, Hardened baselines |
| Unknown senders | DM-capable channels pair unknown senders by default; the operator approves pairing requests | GitHub README |
| Groups | Group access is allowlisted, usually behind a mention | Security |
| Tool execution | Tools run on the host for the main session unless sandboxing is configured; turn on sandboxing (off by default) with agents.defaults.sandbox or per agent | Sandboxing |
| Sandbox backends | Docker (default local backend), Podman, SSH, OpenShell and Crabbox; tool allow and deny policies apply before sandbox rules | Sandboxing |
| Escape hatch | tools.elevated runs exec outside the sandbox; the hardened baseline sets it to disabled | Sandboxing |
| Hardened baseline | Messaging tool profile, exec security "deny" with ask "always", workspace-only file access, agent-to-agent access off, DMs scoped per channel peer | Hardened baselines |
| Audit | A built-in security audit command checks configuration and lists findings with severity and auto-fix support | Security |
| Trust boundary | One trust boundary per gateway; OpenClaw "is not a hostile multi-tenant security boundary", so mixed-trust users need separate gateways and credentials | Security |
The documentation describes sandboxing as "not a perfect security boundary" that "materially limits filesystem and process access". It treats inbound messages and any content the agent reads (web pages, emails, documents, attachments) as untrusted, and recommends a read-only reader agent for untrusted content, keeping web and browser tools off for tool-enabled agents unless needed, and enabling strict inline evaluation checks when interpreters are allowlisted. See the Prompt Injection Guide, the AI Agent Sandboxing and Code Execution Guide and, for enterprise allow, contain or block decisions, the Self-Hosted Personal AI Agent Security Guide.
Data handling
- Where data goes: prompts go to the model provider and chat platforms the operator configures, plus any diagnostics export the operator enables, according to the README.
- Telemetry: a daily update check by default sends the OpenClaw version, operating system, Node.js version, CPU architecture and request surface; settled update outcomes send a bounded report. Anonymous feature statistics (configured channels and providers, plugin inventory, a session-creation count) are off by default. Setting update.checkOnStart to false disables the automatic requests.
- State on disk: the docs say to assume anything under ~/.openclaw/ may contain secrets or private data, including openclaw.json (tokens and allowlists), credentials/ (channel credentials such as WhatsApp), state/openclaw.sqlite (pairing approvals and MCP OAuth tokens) and per-agent databases holding model auth profiles and session transcripts.
- Hardening guidance: directories 700 and files 600, full-disk encryption on the gateway host, and a dedicated OS user if the host is shared. Secrets can be held through SecretRef providers (environment, file, exec or store) rather than plain config.
Model API keys and channel bot tokens held by OpenClaw are non-human identities: scope them to the assistant, rotate them and cap their spend. See the LLMjacking Guide.
Support and lifecycle status
| Model or item | Details | Source |
|---|---|---|
| OpenClaw project | Active; security patch release 2026.11.3 is the newest entry in the changelog index on 15 November 2026 | Changelog |
| Legacy model references | codex/* and openai-codex/* refs are treated as legacy state; the doctor command migrates them to standard openai/* refs | Official provider plugins |
| Security updates | Vulnerabilities are reported through SECURITY.md in the repository; apply security releases promptly | GitHub README |
Practitioner checklist
- Install from the official installer, npm package or desktop builds, and stay on the stable or extended-stable channel for any assistant that holds real credentials.
- Keep the Gateway bound to loopback with token authentication, and follow the exposure runbook before exposing it beyond the host.
- Keep DM pairing and group mention gating on, and run phone-number channels on a separate number from your personal one.
- Turn on sandboxing (off by default), keep tools.elevated disabled, and start from the hardened baseline before re-enabling tools per trusted agent.
- Store model keys and bot tokens through SecretRef providers, scope and rotate them, and set spend limits with each model provider.
- Lock down ~/.openclaw/ permissions, encrypt the host disk, and treat session transcripts as sensitive data in backups.
- Review every skill before installing it from ClawHub, and prefer workspace-scoped skills you control.
- Run the built-in security audit after each configuration change and after every upgrade.
- Decide whether to keep the daily update check and confirm anonymous feature statistics stay off unless you choose to share them.
Standards and references
- OpenClaw Foundation: OpenClaw website (accessed 2026-11-15)
- OpenClaw Foundation: openclaw/openclaw GitHub repository and README (accessed 2026-11-15)
- OpenClaw docs: The Lore of OpenClaw (accessed 2026-11-15)
- OpenClaw: Changelog index and release notes 2026.8.1 and 2026.11.3 (accessed 2026-11-15)
- OpenClaw docs: Security, hardened baselines and prompt injection (accessed 2026-11-15)
- OpenClaw docs: Sandboxing (accessed 2026-11-15)
- OpenClaw docs: Secrets, storage, and logs (accessed 2026-11-15)
- OpenClaw docs: Usage telemetry and update checks (accessed 2026-11-15)
- OpenClaw docs: Model providers and official provider plugins (accessed 2026-11-15)
- OpenClaw docs: Chat channels, platforms and release channels (accessed 2026-11-15)
Related NHI Mgmt Group resources: Personal AI Agents, Self-Hosted Personal AI Agent Security Guide, Agent Skills Security Guide and LLMjacking Guide.