I’ve been running OpenClaw on a dedicated Mac Mini for about two to three weeks now. In that time I’ve burned through roughly $700 in API costs, connected it to Telegram, tried and abandoned iMessage, and watched it autonomously research links and draft blog posts on my behalf.

Some of that is genuinely useful. Some of it is genuinely unsettling.

OpenClaw delivers what ChatGPT and Claude have been promising but can’t actually do on their own. 145,000 GitHub stars in 60 days proves I’m not alone in thinking this. But the gap between “this is incredible” and “this has access to everything” is smaller than you’d think.

Quick Verdict

Aspect	Rating
Overall Score	★★★★☆ (3.8/5)
Best For	Personal automation, link research, content workflows
Pricing	Free (pay only for LLM APIs — budget $200-300/month)
Setup Difficulty	Moderate-High (Docker + patience)
Platform Support	Mac, Linux, Windows (WSL)
Security Risk	High (broad permissions — think carefully)
Memory Persistence	Good in theory, frustrating in practice

Bottom line: The most capable open-source AI agent available, but memory and cron features need constant babysitting. Powerful when it works. Budget more than you expect for API costs.

Get OpenClaw on GitHub →

What Makes OpenClaw Different

OpenClaw doesn’t live in a browser tab. It runs on your machine, connects to your actual messaging platforms, and remembers everything. Not session-by-session memory — persistent, searchable context that spans weeks. At least in theory. The reality needs unpacking.

Created by Peter Steinberger (founder of PSPDFKit), OpenClaw started as Clawdbot, became Moltbot, then settled on OpenClaw after Anthropic’s trademark team got involved. The naming chaos hasn’t slowed adoption — it’s now the fastest-growing AI project in GitHub history.

Here’s what sets it apart: OpenClaw connects AI to your actual digital life. While Claude Code stays in your terminal and ChatGPT lives in its web interface, OpenClaw integrates everywhere you already work. The key word being “integrates” — whether it does that smoothly is another question.

My Setup: Dedicated Mac Mini

I want to be clear about one thing: I did not install this on my personal machine. I set OpenClaw up on a standalone Mac Mini specifically for this purpose. Given the permissions this thing requires — read/write access to messaging platforms, file system access, network access, browser automation — I wasn’t comfortable giving it the keys to my daily driver.

If you’re considering OpenClaw, I’d strongly recommend the same approach. Dedicate a machine or VM to it. The convenience isn’t worth the risk of having an AI agent with broad permissions on the same machine where you do your banking and store your passwords.

The Models I’ve Tested

OpenClaw doesn’t care which AI model powers it. I’ve tried a bunch:

• Claude Opus 4.5 and Opus 4.6 — What I’m running now. Best results by far for nuanced responses and complex research tasks.
• Claude Sonnet 4.5 — Solid, cheaper, but noticeably less capable for the multi-step workflows I need.
• OpenAI GPT-5.2 — Fast and reliable but produces flatter output. Fine for routine tasks.
• Gemini Flash — Cheap and quick. Adequate for simple message handling but falls apart on anything complex.
• Local models — Tried a few. The quality gap is still too wide for anything I’d trust to act on my behalf.

Switching models requires changing one line in the config:

llm:
  provider: anthropic
  model: claude-opus-4-6-20260120
  api_key: ${ANTHROPIC_API_KEY}

I settled on Opus because when you’re giving an AI agent the ability to research, write, and post content for you, model quality matters more than when you’re just chatting. The cost difference is real — that’s a big chunk of my $700 — but the output quality gap makes it worth it.

For comparison of these models, see our Claude vs ChatGPT vs Gemini analysis.

Platform Connections: What Actually Works

OpenClaw connects to platforms through official APIs and bridge protocols. I’ve tested several:

Platform	Connection Type	My Experience
Telegram	Bot API	Excellent — my primary interface
iMessage	Mac-only bridge	Laggy and unreliable on Mac Mini
Slack	Slack App	Works well for monitoring
Email	IMAP/SMTP	Solid for drafting
WhatsApp	WhatsApp Business API	Haven’t tested personally

Telegram is the winner here. It became my go-to interface almost immediately. The Bot API is mature, responses come back fast, and the whole experience feels natural. I send OpenClaw a link, it processes it, and the response shows up right in the chat.

iMessage was a different story. On the Mac Mini, it lagged constantly. Messages would take 30-60 seconds to register, responses were delayed, and the bridge protocol felt fragile. I gave up on it after the first week. If you’re running on a Mac Mini rather than your primary machine, expect iMessage to underperform.

My Killer Use Case: Link Research to Notion

Here’s where OpenClaw earns its keep for me. My primary workflow:

1. I send a link via Telegram — could be a social media post, a blog article, a product page, whatever catches my attention
2. OpenClaw stores it in Notion — automatically categorized and timestamped
3. It researches the link — visits the page, reads the content, pulls context
4. It highlights key components — different treatment for social media vs. websites vs. articles
5. It drafts content — depending on the source, it’ll build out a social media post or a blog draft on the topic

This single workflow saves me hours. Instead of bookmarking things I’ll never revisit, I have an agent that actively processes and contextualizes everything I find interesting. The Notion database becomes a living research library, and the draft content gives me a starting point rather than a blank page.

When it works, it’s the best argument for AI agents I’ve seen. The “when it works” caveat matters though.

Memory, History, and Cron: The Frustrating Parts

Most AI assistants forget everything between sessions. OpenClaw promises to fix this with persistent memory stored in a local database with vector embeddings for semantic search. In practice, you’ll spend a lot of time making this work.

Memory is not easy to get right the first time. I spent days configuring how OpenClaw should remember and recall information. The default settings produced either too much noise (recalling irrelevant context) or too little (forgetting things I explicitly told it to remember). Tuning the similarity thresholds and memory categories took real effort.

History retrieval is inconsistent. Sometimes OpenClaw pulls exactly the right context from two weeks ago. Other times it acts like we’ve never spoken. There’s no clear pattern to when it works and when it doesn’t, which makes it hard to debug.

Cron jobs are the worst offender. Setting up scheduled tasks — daily summaries, periodic research, automated check-ins — sounds straightforward. It isn’t. They break silently. You’ll think you have a working daily digest, and three days later realize it stopped firing. Then you fix it, it works for a week, and something else shifts. Cron tasks need ongoing babysitting even after you think you’ve got them working.

This is probably my biggest frustration with OpenClaw. The features that make it compelling as a long-running agent — memory, history, scheduling — are the same features that need the most hand-holding.

The Browser Problem: Powerful and Scary

One of OpenClaw’s most impressive capabilities is browser access. When it remembers it has this ability, it can navigate websites, fill out forms, pull information, and act on your behalf.

The “when it remembers” part matters. Sometimes OpenClaw will try to answer a question from memory when it should just open a browser and check. Other times it’ll browse proactively and bring back exactly what you need without being asked.

But here’s what keeps me up at night: the browser runs with whatever credentials your machine has. If you’re logged into Gmail, your bank, your CMS, your hosting provider — OpenClaw can access all of it. On my dedicated Mac Mini, I’ve limited what’s logged in. On a personal machine? You’d be giving an AI agent access to your entire digital life.

This isn’t theoretical risk. It’s the actual architecture. The agent has a browser. The browser has your cookies. Think about that before you set this up on your daily machine.

The $700 Reality Check

OpenClaw itself is free and open source (MIT license). Your wallet won’t agree that it’s free.

Component	Cost	My Spend (3 weeks)
OpenClaw	$0	Free
Claude Opus API	~$15/M input, $75/M output	~$600
Other models tested	Varies	~$80
Mac Mini	One-time	Already owned
Electricity	24/7 operation	~$10-15

Total so far: ~$700 in about three weeks.

That’s significantly more than most people expect. The Opus models aren’t cheap, and when you have an agent running 24/7, processing links, researching content, and maintaining memory — tokens add up fast. You could cut costs substantially by using Sonnet or GPT-5.2 for routine tasks and only routing complex work to Opus, but I haven’t optimized that yet.

Compare that to Zapier’s $69/month for basic automation. OpenClaw does more, but it costs more too — especially if you want the best models doing the heavy lifting.

Who Should Use OpenClaw

This Makes Sense For

People who process a lot of links and content. If you’re constantly finding stuff online, researching it, and turning that into content, OpenClaw’s link-to-Notion-to-draft pipeline will change your workflow. It changed mine.

Developers comfortable with Docker and debugging. You will spend time configuring, troubleshooting, and reconfiguring. If that sounds like a fun weekend project rather than a chore, you’ll get a lot out of this.

Privacy advocates wanting AI assistance without cloud lock-in. Run everything on your own hardware with your own models if you want.

People comfortable with AI having broad access. This is the big one. OpenClaw works best when it can reach into your tools — Notion, email, messaging, browser. If the idea of an AI agent with access to your accounts makes you uncomfortable, this isn’t for you. That’s reasonable.

Skip This If

You’re not OK with the security trade-offs. OpenClaw needs broad permissions to be useful. There’s no way around this. If you can’t dedicate a machine to it or aren’t comfortable with the access model, stick with ChatGPT Plus or Claude.

Enterprise users needing compliance guarantees. No SOC 2, HIPAA, or GDPR certifications. Try Microsoft Copilot instead.

Non-technical users wanting plug-and-play. Setup requires Docker, environment variables, API keys, and a willingness to debug. Things will break.

A Word About Privacy: Don’t Do What Everyone Says

I’ve seen people in the OpenClaw Discord and subreddit feeding the agent their entire life story. Health history, financial details, relationship information, daily routines, passwords — everything — because “the more context it has, the better it works.”

Don’t do this.

Treat OpenClaw the same way you’d treat any other AI platform. It’s another instance of Claude or ChatGPT running on your machine. The fact that it’s local doesn’t make it inherently safe — dependencies get compromised, configs get shared, machines get accessed.

Be intentional about what you share. Give it the context it needs for the tasks you want it to do. Don’t hand over your medical records because a Reddit post said it would be helpful. The same information hygiene you’d practice with ChatGPT applies here. Maybe more so, given OpenClaw’s broader access to your systems.

Setup: What You’re Actually In For

1. Clone the repository

   git clone https://github.com/openclaw/openclaw.git
   cd openclaw

2. Configure your environment

   cp .env.example .env
   # Add your API keys and platform credentials

3. Run with Docker

   docker-compose up -d

4. Connect Telegram first (easiest and best integration):

   • Create bot via @BotFather
   • Add token to .env
   • Message your bot to initialize

5. Set up Notion integration (if using the link research workflow):

   • Create a Notion integration
   • Share a database with it
   • Add the token and database ID to config

6. Configure memory and cron carefully Don’t accept defaults. Spend time with the memory configuration. Start conservative and expand from there. For cron jobs, test each one manually before trusting it to run unattended. Then check on them regularly anyway.

7. Budget your first week API costs spike during initial setup — indexing history, testing prompts, training on your style. Budget extra credits for the first week. Then budget more for the second week when you realize the first budget wasn’t enough.

Setup takes 2-3 hours if you know Docker. Budget a weekend if you’re learning as you go. The official documentation is decent but won’t cover every edge case you’ll hit.

My Verdict After Three Weeks

OpenClaw delivers real automation. The link research pipeline alone justifies the setup time for my workflow. When the memory works, when the cron jobs fire, when it remembers it has a browser — it’s the most capable AI agent I’ve used.

But it’s held together with more duct tape than the GitHub stars suggest. Memory needs constant tuning. Cron breaks quietly. iMessage integration isn’t ready for a headless Mac Mini. And the security model requires you to make peace with an AI agent that has real access to real accounts.

I’m keeping it running. The $700 in three weeks stings, but I’ll optimize the model routing to bring that down. The Notion research pipeline saves me enough time that the math works, even at current costs. Just go in with realistic expectations about what “works out of the box” actually means — and don’t install it on your personal machine.

Frequently Asked Questions

Is OpenClaw really free?

The software is free and open source (MIT license). The API costs are not. Running Opus models 24/7 cost me roughly $700 in three weeks. You could cut that significantly with cheaper models, but “free” is misleading if you want good output.

Which model should I use?

I’ve had the best results with Claude Opus 4.5 and 4.6 for complex tasks. GPT-5.2 works fine for routine message handling. Gemini Flash is adequate for simple stuff. Local models weren’t good enough for anything I’d trust to act autonomously.

How much technical knowledge do I need?

More than moderate. You need Docker, environment variables, API keys, and the patience to debug cron jobs that stop working without explanation. If you can follow a GitHub README and troubleshoot on your own, you’ll get there. Budget more time than you think.

Should I install it on my personal machine?

No. Use a dedicated machine or VM. OpenClaw needs broad permissions and browser access. On your personal machine, that means it can access anything you’re logged into. A $500 Mac Mini is worth the separation.

What about the memory features?

They work, but not reliably out of the box. Expect to spend real time tuning similarity thresholds, memory categories, and retrieval settings. Even after you think you’ve got it right, you’ll find cases where it forgets things it shouldn’t or recalls things that aren’t relevant.

Is Telegram really better than iMessage?

For OpenClaw on a Mac Mini, yes, significantly. Telegram’s Bot API is fast and reliable. iMessage through the Mac bridge protocol lagged badly on my setup — 30-60 second delays on message registration. If you’re running on a primary Mac with iMessage already active, results may differ.

How do I keep costs down?

Route simple tasks to cheaper models (Gemini Flash, GPT-5.2) and only send complex research and writing tasks to Opus. I haven’t fully implemented this yet — it’s next on my list. The config supports model routing rules, but getting them right is another configuration project.

Can I trust it to post content on my behalf?

With supervision, yes. I use it to draft content, not publish directly. The quality from Opus is good enough that I’m editing rather than rewriting, but I always review before anything goes live. Full autonomous posting is technically possible. I wouldn’t recommend it yet.

---

Last updated: February 7, 2026. Based on approximately three weeks of daily use on a dedicated Mac Mini. Features verified against OpenClaw v2.1.0.