What people are actually building with OpenClaw (not demos - real use cases)

By Linas Valiukas · March 30, 2026

The OpenClaw subreddit has 200,000 members. The GitHub repo passed 340,000 stars. Jensen Huang called it "the most important release of software, probably ever." So where are the users?

That question shows up in every Hacker News thread about OpenClaw. Someone posts their setup, and half the comments are "cool, but is anyone actually using this in production?" Fair question. We dug through HN threads, Reddit posts, the awesome-openclaw-usecases repo (28,000 stars), and dozens of blog posts and tweets to find out.

Here's what we found. These are real people, most of them identifiable by username, with specific details about what they built, what it cost, and what broke.

The personal assistant builds

This is the most common pattern by far. Someone sets up OpenClaw on a Mac Mini or a VPS, connects it to Telegram or iMessage, and starts texting it tasks. The best setups we found go way beyond "remind me to buy milk."

The "second brain" that actually works

Brandon Wang wrote the most detailed public account we've seen. His OpenClaw instance runs on a Mac Mini and communicates through a private Slack workspace. Every 15 minutes, it scans his iMessages for concrete promises he made - "let me review this tomorrow!" - and silently creates calendar events. When it detects meeting plans forming in a group chat, it drops a hold on his calendar to prevent double-booking.

He has 30+ price alerts running. Not simple "notify me when this drops below $X" alerts - his agent reviews Airbnb listing photos to determine whether a pullout bed is in a separate room. It tracks packages (replaced the Parcel app entirely). It takes photos of his freezer, parses the contents, updates a Notion inventory list, and removes those items from the grocery list.

Restaurant booking: the agent logs into Resy and OpenTable, handles 2FA by reading the code from his texts, cross-references availability with two calendars, and suggests time slots. It even learns cancellation fee policies and adds the cancellation deadline as a separate calendar event.

That's one person's setup. He didn't publish the cost, but based on similar configurations, it's probably $10-20/day in API costs running on Claude Opus.

The $75/week personal staff

A user named arjie on Hacker News runs OpenClaw as a Telegram assistant, mostly on Sonnet, for about $75/week. The use cases are all over the place, and that's sort of the point:

• SOC2 compliance vendor evaluations - "messaged it as things were happening and it made me a nice doc at the end"
• His wife texts the agent to summarize a shared calendar spreadsheet into something she can forward to a friend asking when they'll be in Taiwan
• Bike rain cover alerts - "told it I wanted this functionality and it wrote something and scheduled it"
• Hardware deal hunting - watches specific subreddits for servers with SXM5 boards and Mac Studios with more than 64GB of RAM
• A daily HN digest that filters out culture war posts and only surfaces AI-related stories it thinks he'd find interesting

None of these are impressive on their own. A lot of them could be done with IFTTT or a Python script. The difference is he described all of them in plain English via Telegram and they just... work. No code, no workflow builder, no maintenance (in theory - we'll get to the reality later).

The phone-from-bed builds

A pattern we didn't expect to see so often: people controlling complex technical work from their phones, while doing something else entirely.

@davekiss migrated an entire website from Notion to Astro and deployed it on Cloudflare without opening a laptop. All through Telegram, while watching Netflix. @georgedagg_ had his agent ("BigC") review failed build logs across multiple Railway services, identify the root cause, update configs, redeploy, and confirm success - via voice commands while walking his dog. @CopyKatCapital built and submitted their first iOS app to Apple TestFlight entirely through Telegram messages. Never opened Xcode.

A Hacker News user named bobjordan takes this the furthest. He runs a supervisor agent called "Patch" through Telegram that coordinates multiple Claude Code instances on a 22-core workstation via tmux and Tailscale. He specs work in structured "beads" that agents implement in parallel. His monthly spend: $400 across Claude Code and OpenAI plans. His claim: "This produces better code than I could write after 10 years of focused daily coding myself." His caveat: "I have hand-held and struggled for over a year getting it all setup the way I need it."

The business use cases that save real money

Car negotiation - saved $4,200

@astuyve on X had their OpenClaw agent negotiate with multiple car dealerships simultaneously via email, iMessage, and browser automation. The agent handled the back-and-forth price haggling autonomously. Result: $4,200 saved on the purchase. We've covered a similar case where the other side turned out to be AI too.

Selling a car on NextDoor

Nalinm on HN: "It's actively selling my dad's car on NextDoor. Fielding inquiries and negotiating. Pretty fun." Short on details but the concept is striking - an AI handling the soul-draining back-and-forth of marketplace negotiations on your behalf.

The $26/month CRM

An open-source community project built an agent-first CRM with 40+ REST API endpoints. It processed 4,000+ emails in 48 hours, handled automated prospecting and lead scoring, and connected to WhatsApp, Telegram, Zapier, and Lemlist. Total cost: $26-300/month depending on volume. The enterprise CRM it replaced cost $30,000/year.

Multi-channel customer service for local businesses

Futurist Systems built a unified inbox for WhatsApp, Instagram, Gmail, and Google Reviews with 24/7 auto-responses for restaurants, clinics, and salons. One restaurant went from 4+ hour response times to under 2 minutes, handling 80% of inquiries without a human. The agent detects language (Spanish, English, Ukrainian) and has clear handoff rules for when a human needs to step in.

The surprisingly creative builds

@dreetje's cleaning lady sent a message about needed supplies. The agent accessed 1Password, handled MFA through Beeper, logged into a grocery website, and placed the order. The entire chain - from a human's text message to a completed grocery order - was automated. "IT built all of this, just by chatting to it on the phone."

Block_dagger on HN analyzed their entire iMessage history to find out which friends are the flakiest. Then had the agent introduce itself in a band group chat and suggest cover songs for an upcoming gig based on the conversation history. That's a weird use case. It's also the kind of thing you'd never find in a product roadmap.

@andrewjiang connected OpenClaw to a $35 holographic cube display and described it as "basically a tamagotchi now." @thekitze hooked it up to a Pebble smart ring for one-tap voice commands from their wrist. @dnouri sends music tracks to the agent and gets back GIFs from the performance plus PDF chord sheets. SwitchBot - a real consumer electronics company - shipped the first hardware product with native OpenClaw support: a home hub that controls smart curtains, locks, plugs, and humidifiers through natural language.

The infrastructure and DevOps builds

Jonahss on HN had OpenClaw rescue a dying media server. "I gave OpenClaw ssh credentials and it updated the OS and packages." When it failed to boot, he sent the agent screenshots. It walked him through journal commands, found 1,300 bad sectors on the drive, copied 1.5TB to a newer drive, and restored everything. "I probably would have thrown the whole box out."

The most extreme DevOps setup comes from a user named Nathan who runs an agent called "Reef" with SSH access, 15 active cron jobs, and 24 custom scripts. It polls a Kanban board every 15 minutes, monitors system health hourly, runs code audits every 12 hours, and does security audits weekly. He uses 1Password CLI for credential injection, TruffleHog for secret scanning, and a private Gitea instance for staging. His lesson, learned the hard way: "AI assistants will happily hardcode secrets. They sometimes don't have the same instincts humans do."

Lxgr on HN captures the vibe best: "Being able to chat with somebody that has a working understanding of a Unix environment and can execute tasks like 'figure out why Caddy is crash looping and propose solutions' for a few dollars per month is a dream come true."

The team and multi-agent setups

Maebert on HN runs a team agent that lives in a group chat: "Runs our standups, checks in with everybody EOD on blockers. Already knows what we shipped on Github and Linear so it can focus on the work that's not tracked. Helps with debugging customer issues. Keeps up with twitter and competitors." The part that surprised them: "I'm honestly blown away by the social aspect of it... having an AI team mate is actually fun. Everybody on the team said they'd be sad if we took it away."

Solo founders are running 4+ named agents as a virtual team. @iamtrebuh has Milo (strategy), Josh (coding), Angela (marketing), and Bob (business ops), all managed through Telegram. Some users run 15+ agents across multiple machines with shared memory files - GOALS.md, DECISIONS.md, PROJECT_STATUS.md - and private per-agent directories.

A children's gaming portal at elbebe.co runs an autonomous game development pipeline that produces one educational game or bugfix every 7 minutes. Pure HTML5, conventional commits, feature branches. A solo parent-developer runs the whole thing.

What it actually costs

The range is enormous, and most people underestimate it.

• Light personal assistant usage: $1-2/day ($30-60/month)
• Heavy daily driver on Sonnet: ~$75/week ($300/month)
• Agentic coding on Opus: $1-2/day on light days, up to $110/day when going hard
• Infrastructure alone (VPS/Mac Mini, not counting API costs): $50-100/month

One user put it bluntly: the agent only really works well on Opus-class models. The cheaper models "critically struggle to grep the full array of tools they have available." Another reported their agent with Kimi K2.5 would insist it didn't have calendar access, but if you asked four or five times in a row, it would "discover" the tool.

The honest criticisms

We'd be lying if we said every story was a success. The HN threads are full of people who tried and quit.

Mikenew spent three days doing a deep evaluation. He liked three things: persistent memory across conversations, cross-device messaging (start on desktop, pick up on phone via iMessage), and the breadth of integrations. Then: "However, it's worth stressing how terrible the software actually is. Not a single thing I attempted to do worked correctly." He deleted it and built something simpler.

827a on HN: "OpenClaw feels to me like the promised land of productivity is always over the horizon... tool calling is vastly overblown. It takes forever to get them set up, and that's to get them barely working." And then the reality check: "I oftentimes just don't see the point. I can click the Gmail or Google Calendar app on my phone and get what I need out of those apps in less-than 6 seconds."

The security concerns are real too. One user's agent self-registered on a social network and got banned. Another's revealed family holiday plans to other bots. The 135,000+ exposed instances and the ClawHub malware problem aren't theoretical risks - they're documented incidents.

What separates the builds that work from the ones that don't

After reading through hundreds of reports, a pattern emerges. The people who get real value from OpenClaw share a few traits:

• They use it for things that are annoying but not complex - inbox triage, calendar management, price monitoring, follow-ups. Tasks where "good enough" is good enough.
• They treat it as asynchronous. Text it a task, walk away, check results later. The new /tasks command helps here - you can finally see what's running in the background without checking logs. Trying to have a real-time back-and-forth for urgent work doesn't go well.
• They're comfortable spending time configuring and babysitting. The year-long setup stories aren't exaggerations. The people who benefit most have the technical skill to fix things when they break - which is often.
• They run Opus or Sonnet. Cheaper models struggle with tool selection and multi-step reasoning. You can't cheap out on the LLM and expect it to handle 15 integrations.

The people who bounce off it usually try to use it as a replacement for apps they already use efficiently. If you can check your calendar in 6 seconds on your phone, having an AI do it in 30 seconds via Telegram isn't an upgrade. The value shows up in the tasks you wouldn't bother doing at all - scanning 10,000 emails, watching 30 subreddits for deals, calling five vendors for quotes simultaneously.

Skip the year of setup

Every build in this article required the user to self-host OpenClaw. That means Docker setup, API key management, debugging gateway errors, keeping up with weekly releases, and dealing with security patches. The users who make it work are spending 5+ hours a month on maintenance alone.

TryOpenClaw.ai gives you a managed OpenClaw instance - pre-configured, secured, and maintained - starting at $39/month. Same agent. Same capabilities. Same messaging integrations. No Docker, no SSH, no 3am gateway crashes. You text your agent, it does things. That's the whole product.