Frequently Asked Questions¶

The real questions people actually ask -- ranked by how often they come up on Reddit, GitHub Issues, and community forums.

Last updated: February 14, 2026

Table of Contents¶

Costs & Pricing
Setup & Installation
Security & Safety
Models & Configuration
Features & Capabilities
Comparisons
Technical Deep Dives

Costs & Pricing¶

How Much Does OpenClaw Actually Cost to Run?¶

The #1 question. Software is free (MIT license). You pay for LLM API calls.

Usage Level	Monthly Cost	How
Free	$0	OpenRouter free models, Gemini free tier, or local Ollama
Light personal	$5-15	Kimi K2.5, DeepSeek, or Gemini paid tier
Active personal	$20-50	Smart model routing (cheap + premium mix)
Heavy automation	$50-150+	Multiple agents, always-on, premium models
Unmanaged overnight	$200-700+	What happens without spending caps

The sweet spot: $10-50/month with smart model routing (cheap model for background tasks, premium for interactive).

Can I Run It Completely Free?¶

Yes. Multiple free options: - OpenRouter free models -- Llama 4, DeepSeek R1, Gemini Flash, GPT-OSS 120B (rate limited) - Google AI Studio -- Gemini 2.5 Flash free tier (1,500 req/day) - NVIDIA API -- Free Kimi K2.5 access - Ollama -- Run models locally, $0 forever (need 8GB+ RAM)

Quality tradeoffs exist vs paid models, but it works for learning and light use.

Can I Use My Claude Pro/Max Subscription Instead of API Keys?¶

Technically possible, but you'll likely get banned.

Anthropic actively bans users who route subscription tokens through third-party tools like OpenClaw. Massive ban wave in January 2026. Multiple confirmed bans on Reddit and GitHub. Zero warning before ban.

Safe options: - Anthropic API keys from console.anthropic.com (pay-per-use) -- fully legal - Any non-Anthropic model -- no restrictions - Community advice: "Just use Kimi K2.5 or DeepSeek. Not worth the ban risk."

See Subscriptions & Bans for the full breakdown.

How Do I Control Costs / Set Spending Limits?¶

Set hard cap at API provider -- Anthropic, OpenAI, OpenRouter all support spending limits
Use cheap models for background tasks -- Kimi K2.5 ($0.10/M cached) or Gemini Flash Lite for heartbeats
Restrict cron jobs to business hours -- no 3am tasks
Set concurrency limits -- maxConcurrent: 1 for sub-agents
Monitor weekly -- check provider dashboard for token usage

Setup & Installation¶

How Do I Set It Up? What's the Fastest Path?¶

npm install -g openclaw@latest
openclaw onboard --install-daemon
openclaw gateway
openclaw dashboard    # Opens http://127.0.0.1:18789

10 minutes from zero to working agent. Full guide: Getting Started

Does It Work on Windows?¶

Not natively. You MUST use WSL2 (Windows Subsystem for Linux).

# Install WSL2 first
wsl --install

# Then install inside the Ubuntu environment
# Everything else works the same as Linux

"Never ever try OpenClaw on Windows natively" is common Reddit advice. macOS and Linux work natively.

Do I Need a Mac Mini / Expensive Hardware?¶

No. OpenClaw calls cloud APIs -- it doesn't run LLMs locally. A $4-6/month VPS is enough.

Mac hardware only matters if you ALSO want to run local models via Ollama alongside OpenClaw. For OpenClaw alone, any computer or cheap VPS works.

My Setup Is Stuck / Onboarding Won't Complete¶

Most common causes: 1. Wrong Node.js version -- need 22+ (node --version) 2. Invalid or expired API key 3. Network issues (firewall blocking outbound HTTPS) 4. Missing dependencies

Fix: openclaw doctor --fix catches most issues.

How Do I Connect WhatsApp / Telegram?¶

Telegram (easiest): 1. Find @BotFather on Telegram -> /newbot -> copy token 2. openclaw channel add telegram --token YOUR_TOKEN 3. Message your bot -> enter pairing code

WhatsApp: 1. Open dashboard at http://127.0.0.1:18789 2. Channels -> Add -> WhatsApp -> scan QR code with phone

Warning: WhatsApp links to YOUR number. The agent sees ALL messages. Use a dedicated number.

I Get No Replies in Telegram/WhatsApp. What's Wrong?¶

Check gateway: openclaw gateway status
Check API key: openclaw models status --probe
Check channel config: verify token is correct
WhatsApp: pairing may have expired -- re-scan QR
Run: openclaw doctor --fix

Security & Safety¶

Is OpenClaw Safe? Can the AI Access All My Files?¶

Yes, it has broad access by design. This is the trade-off for autonomous capability.

Real risks: - Full file system access (reads/writes your files) - Shell command execution - Browser control (cookies, sessions) - API key exposure if misconfigured - Prompt injection attacks via malicious content

Mitigations (do these): - Run on a dedicated VPS, not your personal machine - Enable Docker sandbox for tool execution - Use burner accounts -- never connect primary email or banking - Set DM policy to "pairing" (requires approval codes) - Run openclaw security audit --deep before going live

See Getting Started - Security Hardening for the full checklist.

Should I Be Worried About Prompt Injection?¶

Yes. OpenClaw has shell access. A malicious website, email, or message could contain hidden instructions that trick the agent into executing commands.

Protect yourself: - Enable Docker sandbox - Load security instructions into agent memory - Never give the agent production credentials - Use scoped API tokens with minimum permissions - Don't let it browse untrusted websites unsupervised

135,000+ Instances Were Found Exposed on the Internet. How?¶

People bound their gateway to 0.0.0.0 (all network interfaces) instead of 127.0.0.1 (localhost only). This exposed their full OpenClaw setup -- API keys, conversations, credentials -- to anyone who scanned for port 18789.

Fix: Always bind to 127.0.0.1. Use Tailscale for remote access.

Models & Configuration¶

What Model Should I Use?¶

Budget setup (recommended for starting out):

Task	Model	Cost
Default/interactive	Kimi K2.5 or DeepSeek V3.2	$0.10-0.53/M tokens
Heartbeats/background	Gemini 2.5 Flash Lite	Free or $0.50/M
Complex reasoning (when needed)	Claude Sonnet 4.5 or GPT-5	$11-18/M

Free setup:

OpenRouter free models (meta-llama/llama-4-maverick:free, deepseek/deepseek-r1:free, google/gemini-2.5-flash:free) or Google AI Studio free tier.

Premium setup:

Claude Opus 4.6 for maximum quality. But at $30/M tokens, only use for complex tasks.

How Do I Use Different Models for Different Tasks?¶

Configure tiers in openclaw.json:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "openrouter/moonshotai/kimi-k2.5",
        "fallbacks": ["openrouter/google/gemini-2.5-flash:free"]
      },
      "heartbeat": {
        "model": "openrouter/google/gemini-2.5-flash-lite:free"
      }
    }
  }
}

Use /model opus in chat to switch models on the fly for specific tasks.

Can I Run Local Models with Ollama?¶

Yes. Install Ollama, pull a model, point OpenClaw at it:

ollama pull deepseek-r1:14b
ollama serve    # localhost:11434

Configure in openclaw.json:

{
  "agents": {
    "defaults": {
      "model": { "primary": "ollama/deepseek-r1:14b" }
    }
  }
}

Reality check: Local models work but tool-calling support is hit-or-miss. Qwen3 and GLM-4.7 have the best local tool-calling support.

Features & Capabilities¶

What Can OpenClaw Actually DO?¶

Verified real-world uses (from Reddit): - Email triage -- scan inbox, categorize, draft responses (78% time reduction) - Content production -- write blog posts, publish to WordPress, SEO optimization - Calendar management -- schedule, reschedule, send reminders - Browser automation -- research, fill forms, compare prices, negotiate - Code management -- spawn Claude Code/Codex, run tests, create PRs - Social media -- post to Twitter/X, monitor Reddit, engage on LinkedIn - Customer support -- respond to WhatsApp/Telegram inquiries - Business operations -- daily digests, heartbeat monitoring, reporting - Voice calls -- make and receive calls via ElevenLabs + Telnyx/Twilio

What Doesn't Work Well?¶

Setup is harder than advertised -- YouTube makes it look easy, reality is more complex
Costs spiral without monitoring -- agents don't stop when you stop watching
Not enterprise-ready out of the box -- no RBAC, SSO, or audit logging
Security is a real concern -- 341 malicious skills found on ClawHub
Simpler tools often suffice -- n8n or Make.com can handle many automations without AI

Can It Run 24/7 Overnight?¶

Yes, but set spending caps first. Real horror stories: - $18.75/night from a heartbeat asking "Is it daytime yet?" - $200 overnight from "simple scheduled tasks" - $500+/month from unmanaged research swarms

See Getting Started - 24/7 Operation for cost control.

What's a "Heartbeat"? What's a "Cron Job"?¶

Heartbeat: The agent wakes up at regular intervals (default: 30 minutes) and checks for tasks. Configure what it checks in HEARTBEAT.md.

Cron job: A scheduled task at a specific time (e.g., "every weekday at 7am, send me a morning briefing"). Standard cron syntax.

Use cheap models for both -- Gemini Flash Lite, not Claude Opus.

Comparisons¶

OpenClaw vs ChatGPT / Claude¶

	OpenClaw	ChatGPT / Claude
Runs where	Your hardware / VPS	Cloud (Anthropic/OpenAI servers)
Autonomy	Autonomous -- acts on its own 24/7	On-demand -- you ask, it answers
Actions	Can execute: shell, files, browser, email, calendar	Can only suggest actions
Channels	WhatsApp, Telegram, Slack, Discord, etc.	Web/app only
Privacy	Data stays on your machine	Data goes to provider
Cost model	Pay for API tokens	Subscription ($20-200/mo)

OpenClaw vs n8n / Make / Zapier¶

Different paradigms entirely:

	OpenClaw	n8n / Make / Zapier
Logic	AI decides (non-deterministic)	You define workflows (deterministic)
Setup	Natural language instructions	Visual drag-and-drop builder
Best for	Ambiguous, creative, conversational tasks	Repeatable, compliance-heavy workflows
Reliability	Can hallucinate or make unexpected decisions	Does exactly what you configure
Debugging	Hard (AI reasoning is a black box)	Easy (follow the workflow steps)

They complement each other. Many people use n8n for structured automation and OpenClaw for adaptive, intelligent tasks.

What's Clawdbot / Moltbot / OpenClaw?¶

Same project, three names. Created November 2025 as "Clawdbot" (Claude wordplay), renamed "Moltbot" (lobster molting metaphor) after Anthropic's legal team requested a change, then "OpenClaw" on January 30, 2026. Old tutorials referencing Clawdbot or Moltbot still apply.

OpenClaw vs Claude Code¶

	OpenClaw	Claude Code
Scope	Full-life automation (email, calendar, messaging, coding, browsing)	Coding in terminal/IDE
Channels	21 messaging platforms	Terminal only
Proactivity	Cron jobs, heartbeats, runs 24/7	On-demand only
Non-coding	Email, calendar, voice calls, social media	Cannot

They're complementary. OpenClaw can orchestrate Claude Code via its coding-agent skill.

Technical Deep Dives¶

How Does the Agentic Loop Work?¶

Every AI agent follows Thought-Action-Observation (TAO):

1. OBSERVE  -- receive input or results from last action
2. THINK    -- LLM reasons about what to do next
3. ACT      -- execute a tool (shell, browser, API call)
4. OBSERVE  -- capture the result
5. Loop back to THINK until task is done

OpenClaw's Pi Agent Runtime implements this loop. Each iteration costs tokens. Complex tasks may loop 10-50+ times.

See Forks & Agentic Loops for the full technical breakdown.

Where Is My Data Stored?¶

Data	Location
Config	`~/.openclaw/openclaw.json`
Sessions	`~/.openclaw/sessions/`
Memory	`~/.openclaw/memory/<agentId>.sqlite`
Credentials	`~/.openclaw/credentials/`
Skills	`~/.openclaw/workspace/skills/`

Everything is local. But if you use cloud APIs (Claude, GPT, Gemini), your prompts ARE sent to their servers.

Is macOS Based on Linux?¶

No. Common misconception. macOS and Linux are cousins, not parent-child.

Unix (AT&T Bell Labs, 1969)
  |
  +-- BSD (Berkeley, 1977)
  |     +-- NeXTSTEP (Steve Jobs, 1989)
  |           +-- macOS (2001 - present)
  |
  +-- Linux (Linus Torvalds, 1991) -- COMPLETELY SEPARATE

macOS is formally UNIX-certified. Linux is "Unix-like" but not certified. They share zero source code -- only conceptual heritage from original Unix.

Why Is Apple Silicon Good for AI?¶

Unified Memory Architecture (UMA). CPU + GPU share the same memory pool (up to 512GB). No PCIe bottleneck. A 70B model fits in 192GB unified memory directly. Plus: 10-30W power, silent operation, $10-25/year electricity.

See Hardware Guide for the full comparison.

Do I Need Docker for OpenClaw?¶

No. Community consensus (Feb 2026): Docker is unnecessary overhead for most OpenClaw setups. Direct npm install is simpler and avoids gateway token loops.

# All you need:
npm install -g openclaw@latest
openclaw onboard --install-daemon

Docker is only useful for: - Tool execution sandboxing ("sandbox": { "mode": "non-main" }) - Headless Chromium isolation for browser automation on VPS - Multi-agent isolation in enterprise setups

Source: Multiple community voices. See Getting Started.

Can Anyone Actually Make Money with OpenClaw?¶

Yes, but not how you think. The people making real money are:

Selling to other OpenClaw users -- hosting ($49-199/mo), setup services, courses
Operational efficiency -- $35K/yr SaaS replacement, $0.60/report research automation
Crypto automation -- Bittensor mining, Polymarket arbitrage, DeFi yield
Consulting -- Enterprise deployment ($5K-50K per engagement)

What doesn't work: expecting the AI to independently generate revenue from scratch.

"Stop blaming your OpenClaw agent for not making money. You literally can't make $1k yourself, why would an AI you're prompting make $1M?" -- @iwantlambo

See Real-World Usage and Twitter Insights -- Money Debate.

What's the Best Model for OpenClaw?¶

Depends on budget. Community consensus (Feb 2026):

Budget	Primary Model	Heartbeat/Cron	Fallback
Free	Kimi K2.5 (free via OpenRouter)	Gemini Flash (free)	GLM-4.7
$50-100/mo	Kimi K2.5 or GLM-5	Haiku 4.5	Multiple via OpenRouter
$200+/mo	Claude Opus 4.6	Grok 4.1 FR	Kimi K2.5, GLM-5

"Kimi K2.5 is highest used model for OpenClaw" -- @inqusit

See Model Landscape.

Should I Use a Managed Hosting Service?¶

If you're non-technical, yes. Several options exist (Feb 2026):

If you want...	Use
Zero setup, free trial	Ampere.sh ($500 free credits)
Dedicated Hetzner instance	LobsterFarm.ai
Physical hardware, plug & play	ClawBox ($199)
Full control, cheapest	DIY VPS on Hetzner ($4-7/mo)

Creator's warning: Peter Steinberger advises against quick-install services for security learning reasons. But managed services are fine for people who value convenience over education.

See Getting Started -- Managed Hosting.

Does OpenClaw Support MCP (Model Context Protocol)?¶

Not natively (as of Feb 2026). MCP support is a community feature request: - GitHub Issue #4834 -- Native MCP server support - GitHub Issue #8188 -- MCP Client Support

Community workaround: openclaw-mcp-plugin by lunarpulse enables MCP integration via plugin.

OpenClaw has its own tool/skill system that serves a similar purpose. Claude Code has native MCP support built in.

What Is BMAD-METHOD? Should I Use It?¶

BMAD (Breakthrough Method for Agile AI-Driven Development) is a structured framework with 21 specialized AI agents and 50+ workflows. 35.6K GitHub stars.

Use it if: - Complex, multi-phase project (enterprise, regulated, security-critical) - You want predictable AI development process - Your project has 10+ files and multiple services

Skip it if: - Simple bug fix, config change, or quick prototype - You know exactly what you want and can articulate it clearly - Solo project with clear scope

Works with: Both OpenClaw and Claude Code. Native Claude Code integration via npx bmad-method install --tools claude-code.

See Power User Guide -- BMAD-METHOD.

What Are the Kubernetes Deployment Options?¶

openclaw-helm by serhanekicii provides a Helm chart: - App version: 2026.2.12 - Ports: 18789 (Gateway), 9222 (Chromium headless) - Single-instance only (cannot scale horizontally)

Guides: - Metoro.io K8s Guide -- comprehensive deployment + monitoring - Kuberns -- one-click deployment - DigitalOcean -- 1-click deploy with v2026.1.24-1

Alternative PaaS: Railway, Dokploy, Coolify for simpler container deployments.