What is OpenClaw?
OpenClaw is an open-source personal assistant built around one local Gateway, the control plane for sessions, tools, events and channel connections. You talk to it through WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Teams and more than 20 others; the official catalog lists 33 channels in all. It also has a Control UI, a CLI, a TUI and companion apps for macOS, iOS, Android, Windows and Linux. Models and agent harnesses, including Claude, Codex and local models, are swappable plugins. Skills come from the ClawHub registry.
It is the most-starred project in this section, at about 394,700 GitHub stars for a repo created in November 2025. The npm package openclaw had about 3.12 million downloads in the week to 17 November 2026. It is MIT-licensed TypeScript, stewarded by the OpenClaw Foundation (a 501(c)(3)), with no paid tier, hosted service or token. Releases are date-versioned and frequent; v2026.11.5 shipped on 17 November 2026.
Is OpenClaw worth it?
OpenClaw is the reference point for always-on personal agents. It reaches more chat platforms than its rivals, has native device apps and a skills registry, and has an install base measured in millions of weekly npm downloads. One Gateway on a laptop or server gives the same assistant to your WhatsApp, Slack and phone, with memory and credentials kept on your own hardware. Memory is plain Markdown in the agent's workspace: a curated MEMORY.md and USER.md loaded at the start of each session, plus dated daily notes that a background "dreaming" sweep distils into MEMORY.md. The default engine indexes those files in a per-agent SQLite database with keyword, vector and hybrid search, though vectors come from OpenAI embeddings unless you choose another provider. The project says plainly that by default it phones home only for a daily version check, with anonymous feature statistics opt-in.
The flip side of that reach is power. Sandboxing is off by default, and host exec approvals default to full access with no prompts, so an assistant reachable from a chat app can run commands with your user's rights. Switched on, the sandbox runs tools in a Docker container with no network by default, with Podman, SSH, OpenShell and Crabbox as alternative backends; the non-main mode sandboxes every group and channel session while your own main session stays on the host. Approvals can be tightened to an allowlist or a prompt for every command, and a prompt nobody answers is denied. Inbound messages are untrusted, and DM channels pair unknown senders by default, but every message the agent reads, every web page and every third-party skill is a possible prompt-injection route into that shell. Credentials get more care: config fields can point to 1Password, Vault, Bitwarden or sops instead of holding a key, and a protected secret kind keeps model-provider keys out of the agent's view and substitutes them at the network edge. The masking has limits the docs state plainly: log tails, diagnostics and exports hide sensitive values, but session transcripts and tool output keep credentials in full.
The security record reflects how much surface there is. The repository has published 722 GitHub security advisories: 14 rated critical, 249 high, 390 medium and 69 low, covering things like media-policy bypasses and credentials exposed in logs. Fixing them in public is to the project's credit. The count still means an install left unpatched for a month is probably exposed to something.
The pace cuts both ways. Releases land every few days, and the issue and pull-request queue stands at about 8,300, so features and fixes arrive fast while configuration and plugin behaviour shift underneath you. Compared with nanobot and Hermes Agent, OpenClaw offers more channels, apps and community skills, at the cost of more moving parts to secure and keep current.
Where it earns its place
A freelance developer who lives in Telegram and WhatsApp
Ask for reminders, file lookups and quick scripts from the phone while away from the desk
Install with the script, run openclaw onboard --install-daemon, and pair the phone's chat accounts. Software is free; model use is the bill, and sandboxing should be on before the agent can touch the home directory from chat.
A five-person agency sharing one assistant in Slack
Have a team assistant that drafts client updates and pulls files from a shared drive
Run the Gateway on a small server as a team deployment, connect Slack, and restrict tools with the sandboxing guide and pairing approvals. Plan time each week to apply releases, given the advisory count.
A security engineer asked to approve OpenClaw on staff laptops
Decide what configuration makes it acceptable
Read the security guide and exposure runbook, require sandboxing in all mode, set exec approvals to allowlist, move keys behind SecretRefs (secrets audit finds plaintext ones), disable unneeded channels and skills, set update.checkOnStart according to policy, and track advisories. The software costs nothing; the review and patching time is the real cost.
Before you commit
- Setup
- curl -fsSL https://openclaw.ai/install.sh | bash (or the PowerShell script), which provisions Node and starts onboarding; or npm install -g openclaw@latest on Node 24.22.1+ or 26.12.1+. Onboarding checks model access, creates the workspace and configures the Gateway; then open the Control UI. Without a terminal: OpenClaw Easy, a separate Apache-2.0 desktop app for macOS and Windows from another maintainer, bundles a vendored copy of OpenClaw, installs Ollama and defaults to the local llama3.2:3b model; its WhatsApp setup links your account by QR code as a WhatsApp Web device, so OpenClaw's WhatsApp docs recommend a separate number.
- Integrations
- WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Teams and the rest of 33 listed channels; 1Password, Vault, Bitwarden and sops for secrets; apps for macOS, iOS, Android, Windows and Linux; hosted and local model providers; tools, skills, plugins and the ClawHub registry.
- Data and privacy
- State, credentials and memory (Markdown files indexed in a per-agent SQLite database) stay on your hardware; prompts go to the model and chat platforms you configure, and vector search uses OpenAI embeddings by default. The secret store is SQLite with 0600 permissions, not an HSM. A daily version check is on by default and feature stats are opt-in. 722 GitHub security advisories have been published.
- Cost at scale
- No paid edition. Cost is model usage plus a machine to run the Gateway around the clock, from an existing laptop to a small cloud VM for a team.
- Limits
- Sandboxing is off and host exec approvals default to full with no prompts. Logs are masked but transcripts and tool results keep credentials as written. Plain npm install at the repo root is unsupported for source builds (pnpm workspace). Rapid releases mean frequent configuration changes.
What’s good
- 33 chat channels plus native desktop and mobile apps
- Models and harnesses swap as plugins, including local models
- Memory is plain Markdown on your hardware, searched through SQLite
- Foundation-stewarded, MIT, no paid tier
What’s not
- Sandboxing is off and exec approvals allow everything by default
- 722 published security advisories, 14 of them critical
- Transcripts and tool output keep credentials unmasked
- About 8,300 open issues and pull requests
Best for
Technical users and small teams who want one assistant across their chat apps and devices and will run it sandboxed and patched.
OpenClaw alternatives
OpenClaw: common questions
- Is OpenClaw worth it?
- OpenClaw is the personal agent everyone else is measured against, and it hands a model your shell by default, so it rewards owners who read the security guide first. We rate it 4.3 out of 5.
- How much does OpenClaw cost?
- Free and open source (MIT). No paid tier; you pay your model provider.
- What are the downsides of OpenClaw?
- Sandboxing is off and exec approvals allow everything by default; 722 published security advisories, 14 of them critical; transcripts and tool output keep credentials unmasked; about 8,300 open issues and pull requests.
- Who is OpenClaw best for?
- Technical users and small teams who want one assistant across their chat apps and devices and will run it sandboxed and patched.
- How much does OpenClaw cost for a team or in production?
- No paid edition. Cost is model usage plus a machine to run the Gateway around the clock, from an existing laptop to a small cloud VM for a team.
- Is OpenClaw safe for company data?
- State, credentials and memory (Markdown files indexed in a per-agent SQLite database) stay on your hardware; prompts go to the model and chat platforms you configure, and vector search uses OpenAI embeddings by default. The secret store is SQLite with 0600 permissions, not an HSM. A daily version check is on by default and feature stats are opt-in. 722 GitHub security advisories have been published.
- What are the best OpenClaw alternatives?
- The closest alternatives we have reviewed are Hermes Agent, nanobot, OpenBot and goose.
Answers checked Nov 18, 2026.
