Skip to main content

Frequently Asked Questions

Short answers with links. The official FAQ is at docs.openclaw.ai/help/faq; this page adds our judgement and points to the deeper pages on this site.

General​

What is OpenClaw?​

OpenClaw is a free, open-source AI agent that you run yourself. A long-running Gateway connects LLMs to your files, shell, browser and messaging apps. It has about 392k GitHub stars as of 2026-10-18 and is stewarded by the OpenClaw Foundation.

Is OpenClaw free?​

OpenClaw itself is free (MIT license). You pay for LLM API usage if you use cloud providers. Reported bills range from a few dollars a month to thousands; see How much does it cost to run?. Running local models removes API costs.

Why was it renamed twice?​

  1. Clawdbot (Nov 2025): original name
  2. Moltbot (Jan 27, 2026): renamed after Anthropic asked for a name change ("Clawd" was too close to "Claude")
  3. OpenClaw (Jan 30, 2026): renamed again because "Moltbot never quite rolled off the tongue"

Upstream tells the story in Lore.

Who created OpenClaw?​

Peter Steinberger, an Austrian developer and founder of PSPDFKit. Per the upstream FAQ, he now works at OpenAI; OpenAI is a donor to the Foundation, not an owner of the project (What is OpenClaw?).

What is the OpenClaw Foundation?​

The OpenClaw Foundation is an independent 501(c)(3) that develops OpenClaw in the open, employs the core team and signs releases. Its donor lists differ by source. The README at v2026.10.4 names Amazon, Lobster Computer Company, Offline Holdings, OpenAI, Red Hat and the University of Michigan; the Foundation's donors page (checked 2026-10-18) shows Microsoft, NVIDIA and OpenAI as partners alongside Offline Holdings, Red Hat and the University of Michigan, and does not list Amazon or Lobster Computer Company. Both say donors do not own the project; the README says "none of them own or direct the project". The code stays MIT-licensed.

What is Moltbook?​

Moltbook is a separate social network (created by Matt Schlicht) where AI agents post, comment and vote. It grew to 1.6 million registered agents, most of them bots. It suffered a major database breach in January 2026. It is not part of OpenClaw.

What version is current?​

As of 2026-10-18 the latest stable release is v2026.10.4 (published 2026-10-16). OpenClaw uses calendar versions YYYY.M.N, where N is a release counter within the month, not a day. There are four update channels: stable, extended-stable, beta and dev (development channels). Check yours with openclaw --version.

On 2026-10-18 npm's tags are latest 2026.10.4, extended-stable 2026.9.33 (a trailing, LTS-like line: end of September plus critical security fixes) and beta 2026.10.10-beta.1. Version numbers are not in time order: 2026.10.10-beta.1 was published on 2026-10-15, a few hours before 2026.10.4, because the 10.1 to 10.4 releases come from a maintenance branch while new features ship on the 10.10 beta line. openclaw update --channel beta installs 2026.10.10-beta.1 today. This site documents 2026.10.4 and will move when npm latest does.

The v2026.6.11 regression where tools returned empty output after the first call in a turn (issue #98528) is closed; update to the current release.

Security​

Is OpenClaw safe to use?​

Only if you configure it carefully. It has real access to your machine, and it has had:

Follow Security Hardening before deploying.

Can someone hack me through OpenClaw?​

Yes, if you:

  • Expose the gateway to a network (gateway.bind: "lan" listens on 0.0.0.0) without a strong token and a firewall
  • Install unreviewed skills or plugins
  • Run an outdated version with published advisories
  • Let anyone message the agent (open DM or group policies)

Even loopback-only instances were reachable through the browser before CVE-2026-25253 was fixed. Run openclaw security audit to check your setup.

Does OpenClaw send my data to the cloud?​

  • Config, credentials, sessions and workspace files are stored locally under ~/.openclaw/.
  • LLM prompts go to your chosen model provider, and they include whatever the agent loads into context: messages, tool output, and memory such as MEMORY.md.
  • Local models: prompts stay on your machine.
  • The Foundation receives a daily update check (version, OS, Node version, CPU architecture) by default. Set update.checkOnStart: false to turn it off (telemetry).

What are workspace .env security restrictions?​

Workspace .env files cannot override OPENCLAW_* variables or provider credential keys. This fail-closed design stops a malicious workspace from hijacking credentials. See Environment Variables.

Cost​

How much does it cost to run?​

It depends almost entirely on the model, how often the agent wakes up, and how long sessions grow. Always-on automation with a frontier model is where bills grow fastest. Cost Management has worked estimates.

The main cost drivers are context accumulation (long sessions), the heartbeat (it runs even when idle), and autonomous multi-step tool use. openclaw gateway usage-cost shows what you have actually spent.

How do I reduce costs?​

  1. Use a cheaper or local model for the heartbeat (agents.defaults.heartbeat.model)
  2. Run the heartbeat less often, for example agents.defaults.heartbeat.every: "60m" ("0m" turns the recurring heartbeat off)
  3. Limit it to waking hours with agents.defaults.heartbeat.activeHours
  4. Set agents.defaults.heartbeat.isolatedSession: true so each run starts without the conversation history, and lightContext: true to skip workspace bootstrap files
  5. Use local models for $0 API cost
  6. Keep sessions short, or /compact them
  7. Route sub-agents to cheaper models

See Heartbeat, Cost Management and Performance Tuning.

Can I use it without an API key?​

Yes, with local models via Ollama, LM Studio or vLLM. No API key or internet connection is needed for core operation.

Which LLM should I use?​

Your priorityRecommendation (model ref)
Best qualityanthropic/claude-opus-5-5
Best balanceanthropic/claude-sonnet-5-5
Cheap and fastanthropic/claude-haiku-5-5 or deepseek/deepseek-flash
Free (local)an Ollama model such as ollama/qwen3:14b

A fresh Anthropic setup picks anthropic/claude-opus-5-5. Prices change often; check the provider's pricing page. See the Model Selection Guide and the upstream provider pages.

Can I use multiple models at once?​

Yes. Set a default model, a cheaper heartbeat model, and per-agent overrides under agents.entries.<id>:

~/.openclaw/openclaw.json
{
  agents: {
    ownership: "explicit",
    defaults: {
      model: { primary: "anthropic/claude-sonnet-5-5" },
      heartbeat: { every: "60m", model: "anthropic/claude-haiku-5-5" },
    },
    entries: {
      main: {},
      researcher: { model: "anthropic/claude-opus-5-5" },
      monitor: { model: "ollama/qwen3:14b" },
    },
  },
}

With more than one entry, ownership: "explicit" is required; route channels to agents with bindings (per-agent entries and multi-agent routing). The older agents.list array is retired; openclaw doctor --fix migrates it. For a single agent, openclaw models set anthropic/claude-sonnet-5-5 is enough. See Model Selection.

Technical​

Does it work on Windows?​

Yes. Upstream recommends the native Windows Hub app (openclaw/openclaw-windows-node) for Windows 10 20H2+ and Windows 11. The native CLI and Gateway (PowerShell installer) and WSL2 are the alternatives (Windows).

Does it work offline?​

Yes, with local models. The gateway, memory, skills and file operations work without internet. Cloud LLM calls and chat channels (WhatsApp, Telegram, etc.) need connectivity.

Does it work on Raspberry Pi or ARM?​

Yes. Upstream has an official Raspberry Pi guide: a Pi 5 with 4-8 GB is recommended, a Pi 4 with 2 GB works with swap. Run models in the cloud or on another machine. For even lighter alternatives, see Lightweight Variants.

Can I run it on a cloud server?​

Yes, on any VPS or one of the hosting providers. Keep the gateway on loopback (gateway.bind: "loopback", the default) and reach it through an SSH tunnel, Tailscale or a VPN. See Keep the Gateway on loopback and Deployment Options.

Can I run multiple instances?​

Yes. Use a separate profile for each instance, which gives each one its own config, state and workspace, and run each on its own port:

# Instance 1 (default profile)
openclaw gateway --port 18789

# Instance 2 (a second profile)
openclaw --profile personal setup
openclaw --profile personal gateway --port 19789

See Multiple gateways.

What are the minimum system requirements?​

ComponentMinimumRecommended
OSmacOS, Linux, Windows (native Windows Hub or WSL2)macOS or Linux
Node.js24.16+ or 26.1+26
RAM2 GB (gateway only)4 GB+
Disk500 MB2 GB+

Node 22, 23 and 25 are not supported (Node.js). For local models, add 8-80 GB VRAM depending on model size. See Introduction.

Why is my gateway using so much memory?​

A memory regression in v2026.4.9 raised RSS from about 600 MB to about 946 MB on constrained systems (issue #63526, closed). If memory is still high:

  1. Update to the current release (openclaw update)
  2. /compact long sessions, or start new ones
  3. Restart the gateway (openclaw gateway restart)
  4. Check openclaw gateway stability for a recorded diagnostic

Updates & Migration​

How do I update safely?​

Back up, preview, then update:

openclaw backup create --verify
openclaw update --dry-run
openclaw update
openclaw doctor

If the new version fails validation, the old Gateway keeps serving. Automatic rollback after activation works only when no database schema migration ran, so the backup is your real safety net (Rollback and recovery). If you skipped many releases, run openclaw doctor --fix to migrate retired config keys and files.

Notable breaking updates from 2026:

VersionIssueWhat happened
v2026.5.7Config wipe bug (#80077)The update wiped openclaw.json and deleted credentials; fixed in v2026.5.12
v2026.8.1"OpenClaw 2.0"Sessions moved to SQLite, HEARTBEAT.md retired, cron renamed Automations, many retired config keys refused
v2026.9.3Node runtimeNode 22 dropped; Node 24.16+ or 26.1+ required

How do I recover from the v2026.5.7 config wipe?​

The v2026.5.7 update bug wiped openclaw.json and the credentials/ directory. The built-in .bak files captured the already-corrupted state, so they can't help.

If you have a backup archive: restore it to a staging directory, check it, then copy the files back:

openclaw backup verify ./my-openclaw-backup.tar.gz
openclaw backup restore ./my-openclaw-backup.tar.gz --target ./restored-openclaw

If you don't have a backup:

# Re-run onboarding to recreate config
openclaw onboard

# Re-link channels
openclaw channels login --channel whatsapp
openclaw channels add --channel telegram

Fixed in v2026.5.12 (PR #80257). Update to the current release after restoring.

How do I back up my config?​

Use the built-in backup command. It covers config, credentials, sessions and workspaces (including SOUL.md and MEMORY.md):

openclaw backup create --verify
openclaw backup create --only-config

openclaw backup enable schedules recurring offsite or Git backups. See openclaw backup.

Channels​

Will WhatsApp ban me for using OpenClaw?​

Possibly. The WhatsApp channel uses the unofficial Baileys library (a reverse-engineered protocol), not the official Business API, and WhatsApp detects and bans automated accounts. Risk factors:

  • Sending many messages quickly
  • Automated replies to unknown contacts
  • Running from a VPS IP
  • Using a new number without history

Mitigation: use a number with history, and restrict who can reach the agent with channels.whatsapp.dmPolicy ("pairing" or "allowlist") and allowFrom (WhatsApp). Some users run for months without issues; others are banned within days.

Can I use OpenClaw with email?​

Yes, but email is an automation trigger, not a chat channel. Use the Gmail Pub/Sub trigger or the IMAP trigger plugin. Upstream recommends a restricted, sandboxed reader agent for untrusted mail. Sending email gives the agent a lot of reach; read Exec approvals and tool policy first.

What is Cloudflare Moltworker?​

Moltworker runs OpenClaw in a Cloudflare Sandbox on Workers. Its README calls it an "experimental… proof of concept… not officially supported" that "may break without notice". The README estimates about $34.50/month running 24/7 (the $5 Workers Paid plan plus container compute). Last pushed 2026-05-09.

Ecosystem​

What is ClawHub?​

ClawHub is OpenClaw's skill and plugin registry. It was the target of a malicious skill campaign in February 2026. Since 2026-10-13 new releases stay unpublished until TruffleHog and ClawScan pass, and openclaw skills verify @owner/<slug> fetches a release's audit verdict. A pass is a signal, not a guarantee, so review skills yourself (Skill Verification).

What is Lobster?​

Lobster is OpenClaw's official workflow shell: a typed, local-first macro engine for composable automation pipelines. About 1.3k stars as of 2026-10-18, MIT licensed. See the Lobster Workflows Guide.

What is Clawery?​

clawery.com markets an enterprise AI assistant service ("Your Personal Assistant for Work"). Its site no longer mentions OpenClaw. It is not part of the open-source project.

What are the lightweight alternatives?​

Independent projects, not forks (stars as of 2026-10-18):

ProjectStarsDescription
Nanobot49kSmall Python agent from HKU
ZeroClaw33kRust agent runtime
NanoClaw31kSmall TypeScript agent that runs in containers
PicoClaw30kGo agent for small devices
MimiClaw5.8kRuns on an ESP32-S3 microcontroller

See Lightweight Variants and Ecosystem.

Can I contribute?​

Yes. See Contributing to OpenClaw and Contributing to these docs.

Troubleshooting Quick Reference​

Start with openclaw status --all, openclaw doctor and openclaw logs --follow, or openclaw triage --non-interactive for a sanitized diagnosis (General troubleshooting).

ProblemQuick fix
another gateway instance is already listening / EADDRINUSEAnother process holds the port. openclaw gateway status, then openclaw gateway stop or pick another --port
refusing to bind gateway ... without authA non-loopback bind needs a token. Set gateway.bind back to loopback, or configure gateway.auth (hardening)
Gateway start blocked: set gateway.mode=localRun openclaw config set gateway.mode local, or openclaw doctor --fix if the config was clobbered
Gateway reachable from the networkopenclaw config set gateway.bind loopback, openclaw gateway restart, then openclaw security audit (Keep the Gateway on loopback)
Provider auth errors (401)Check the saved profile with openclaw models auth list; sign in again with openclaw models auth login
Heartbeat costing too muchSet agents.defaults.heartbeat.model to a cheap or local model and lengthen every
Agent won't run commandsopenclaw approvals get shows the effective exec policy. Approve the pending request (openclaw approvals pending, then openclaw approvals resolve <id> allow-once), or allowlist the command (exec approvals)
Sandbox blocking execopenclaw sandbox explain shows what is sandboxed. Adjust the sandbox's workspace access; don't switch the sandbox off
Config wiped after updateRestore from backup; see recovery steps above
High memory usageUpdate; /compact long sessions; openclaw gateway restart
WhatsApp disconnectedRe-link with openclaw channels login --channel whatsapp
Channel messages not arrivingopenclaw channels status --probe; check the channel's dmPolicy and allowFrom
macOS gateway stops overnightPower Nap / Maintenance Sleep; see upstream gateway service troubleshooting
warning

Do not "fix" approval prompts by setting tools.exec.mode: "full" or by turning the sandbox off. Both give the agent unreviewed shell access on your host. See Exec approvals and tool policy and Sandboxing.

For detailed troubleshooting, see Troubleshooting.

See Also​