We Tested 5 Ways to Run an OpenClaw Agent. Here's What Actually Works.
207K GitHub stars means everyone wants to try it. But the deployment options have multiplied faster than the documentation. We spent two weeks running every major option — self-hosted to one-click cloud — so you know what you're getting before you commit.
Part 10 of the Agent Infrastructure Series
OpenClaw has 207,000 GitHub stars. That number keeps going up by roughly 2,000 a day.
That means a lot of people are watching. And every week, a new “deploy OpenClaw in one click” service appears. There’s EasyClaw — actually five different products using that name. There’s InstaClaw. There’s OpenClaw Cloud, the official hosted version. There are two completely different GitHub repositories both called TinyClaw. And there’s the original: self-hosting on your own hardware, which is what the project was designed for in the first place.
The problem isn’t that there are too many options. It’s that most reviews don’t tell you what the experience actually feels like once you’re past the landing page.
So we tested them all. Two weeks, five setups, real tasks. This is what we found.
What We Tested (and How)
Five configurations:
- Self-hosted Mac Mini M4 — the canonical setup
- Self-hosted VPS — Linux cloud server, SSH-only
- TinyClaw — AgentPuter’s cloud runtime, one-click deployment
- OpenClaw Cloud — the official hosted version, $9.99/month
- EasyClaw.ai — three-tier managed hosting ($5–$49/month)
For each setup, we measured:
- Time to first message (from zero to talking to your agent)
- Reliability over 7 days (uptime, dropped messages, restart behavior)
- Feature completeness (what actually works vs. what requires extra configuration)
- Real task performance (email triage, calendar management, web research)
- Monthly cost at moderate use (~2 hours of active use daily)
The model was Claude Opus 4.6 throughout, except where a different model was the only option. All tests used Telegram as the primary channel because it’s the most stable across all platforms.
Option 1: Self-Hosted Mac Mini M4
Time to first message: 22 minutes
7-day uptime: 100%
Monthly cost: ~$35 (API) + $0 amortized hardware
This is the baseline, and it sets a high bar.
The installation is a single command — openclaw onboard --install-daemon — and the wizard walks you through every step. The bottleneck isn’t software; it’s API key management and Telegram bot setup, which takes about 10 minutes on its own if you’ve never done it before.
Once running, the Mac Mini setup beats every cloud option at a few specific things. iMessage integration via BlueBubbles works, and nothing else can match it. Voice Wake lets you speak to your agent hands-free through the native macOS app — unavailable on any cloud deployment. Local filesystem access means your agent can work with files that never leave your machine.
The SOUL.md file is where this starts to feel different from everything else. It’s not a system prompt. It’s a persistent identity file that your agent reads at startup every day. After a week of adding to it, the agent knew my schedule, my communication style, my preferences for email responses, what calendar events I consider optional. That accumulation of context doesn’t exist at the session level — it compounds.
One annoyance: iMessage authentication via BlueBubbles is fiddly and needed re-login every 2–3 weeks in testing. The openclaw doctor command catches this before you lose a day of messages, but you have to actually run it.
If you have a Mac Mini or are willing to buy one, this is the answer. Nothing else comes close on the feature ceiling.
Option 2: Self-Hosted VPS (Linux)
Time to first message: 28 minutes
7-day uptime: 100%
Monthly cost: ~$35 (API) + $6 (VPS)
For Linux users and developers who want root access, this is equivalent to the Mac Mini setup minus Voice Wake and iMessage. The installation is identical; the difference is that you’re maintaining a server rather than a desktop.
The advantage is remote access by design. Your agent is reachable whether or not your laptop is open. Tailscale Serve or Funnel (the officially recommended approach) makes this secure without exposing the Gateway port directly.
The drawback is DevOps overhead. When your agent stops responding at 2 AM, you’re SSH-ing into a server to check logs. If you’re comfortable with that, this is a solid option. If you’re not, skip to the hosted options.
One thing worth flagging: Docker sandboxing — which OpenClaw recommends for isolating the agent’s shell access — adds noticeable latency on VPS setups. Chat responses are fine. Tasks involving file operations or web browsing took 15–30 seconds longer than the Mac Mini equivalent in testing. If you’re not worried about security isolation from the agent itself, running without Docker is faster.
Same ceiling as Mac Mini, more flexibility, more maintenance. Best for developers who already manage servers and aren’t looking for another new thing to learn.
Option 3: TinyClaw
Time to first message: 48 seconds
7-day uptime: ~99.9%
Monthly cost: $49.99/month (model included)
TinyClaw (tinyclaw.dev) is AgentPuter’s cloud runtime for OpenClaw. Each user gets an isolated container rather than a slice of shared compute. In practice, this means your agent’s response time doesn’t crater when a hundred other users send messages at once.
Onboarding was the fastest of anything we tested — Google sign-in, select Telegram, paste a bot token, and the first message came back in 48 seconds. Over seven days, there was one maintenance window of about 4 minutes that was communicated ahead of time.
Response times averaged 6–8 seconds for standard queries, and stayed there. That’s the practical advantage of isolated containers over shared compute: not faster in ideal conditions, but consistent when conditions aren’t ideal. Your home Mac Mini’s internet connection is a variable; a cloud container’s is not.
Skills installation was solid. Of the 12 Skills we tested from ClawHub, 11 installed without issues. The one failure — a Spotify integration that requires OAuth redirect configuration — would need manual setup on any platform.
No Voice Wake, no iMessage, no local filesystem. These are structural cloud deployment limitations, not specific to TinyClaw. If you need any of those, self-hosting is the only path.
The AgentPuter backstory matters if you’re thinking about growing beyond one agent. The same infrastructure supports multiple OpenClaw instances under one account — different agents for different contexts, or shared access across team members. That’s not in the basic plan, but it’s there if your needs expand.
Option 4: OpenClaw Cloud
Time to first message: 2 minutes 10 seconds
7-day uptime: ~98.4%
Monthly cost: $0 (free tier, 14 compute-days/month) or $9.99/month Pro
OpenClaw Cloud (open.claw.cloud) is the officially endorsed hosted version. The main differentiator is the free tier: 14 days of compute time per month, powered by Kimi K2.5. Note that when we tested, the $9.99/month Pro plan was still marked “Coming Soon” — so some of this section is based on the free tier, which is what was actually available.
The free tier is more than a demo. You get pre-installed Skills, Telegram integration, persistent memory, browser automation, and your own dedicated virtual machine. When your 14 free compute-days run out, the VM powers off but your data stays intact. Power it back on when you have more free time, or when the Pro plan launches.
The Kimi K2.5 model is worth a comment. For most practical tasks — summarizing emails, drafting replies, calendar management — it held up fine in testing. The gap with Claude Opus 4.6 showed up in complex, multi-step reasoning: “research three competitors, compare their pricing, and draft a recommendation” came out noticeably more coherent with Claude. That said, for anyone running lightweight workflows, Kimi K2.5’s price-to-performance is hard to argue with.
Onboarding takes slightly longer than TinyClaw because you’re configuring through a web dashboard rather than a wizard. Getting Telegram connected took about 2 minutes. Skills management is fully exposed, which is good if you know what you want and confusing if you don’t.
Cheapest path to a real OpenClaw experience. Best choice if you want to try before committing anything.
Option 5: EasyClaw.ai
Time to first message: 4 minutes
7-day uptime: 99.1%
Monthly cost: $19–$49 depending on tier
EasyClaw.ai’s distinguishing feature is browser-based desktop access — a full view of your agent’s container environment in a browser tab. You can log into accounts that require browser authentication, install tools, modify config files. Things managed platforms don’t normally let you touch.
We used this to set up a Gmail integration that requires authenticating through a browser flow — straightforward on EasyClaw, impossible on platforms without it.
The pricing tiers run $5 (developer, bring your own VPS), $19 (managed), and $49 (fully managed with higher-tier support). At $49/month, you’re close to the point where a self-managed VPS makes more financial sense unless you genuinely want the hands-off maintenance.
One naming note: “EasyClaw” is being used by at least five different products from different developers — easyclaw.ai, easyclaw.app, easyclaw.pro, and a couple more. The one we tested is easyclaw.ai. At least one of the other namealike services raised security questions in the community in early February 2026. Verify you’re at the right URL before entering credentials anywhere.
The Open-Source Variant Landscape
Beyond hosted services, the open-source derivatives are worth knowing about, though they’re aimed at a narrower audience.
TinyClaw (jlia0 version) is a reimplementation of the OpenClaw concept in roughly 400 lines of shell script — built with Claude Code and tmux rather than Node.js and a full framework. It runs multiple isolated agents in parallel across Discord, WhatsApp, and Telegram, coordinating through file-based message queues that prevent race conditions. About 2,000 GitHub stars as of February 2026, with the codebase small enough to read in an afternoon. If you want to understand what’s actually happening under the hood, or need to embed an agent into a constrained environment, it’s the most direct way in.
TinyClaw (warengonzaga version) is a different project entirely — a TypeScript rewrite that positions itself explicitly as “a completely independent product and an alternative to OpenClaw.” It emphasizes plugin architecture, self-improving memory, and smart query routing to reduce LLM costs. Less focused on task execution, more focused on “personal AI companion.” Worth watching, but early.
Neither of these competes directly with hosted services. They’re tools for developers who want to understand and modify the underlying mechanism.
Comparison Table
| Mac Mini | VPS | TinyClaw | OpenClaw Cloud | EasyClaw.ai | |
|---|---|---|---|---|---|
| Setup time | 22 min | 28 min | 48 sec | 2 min | 4 min |
| Monthly cost | $35 API | $41 | $49.99 | $0–$10 | $19–$49 |
| 7-day uptime | 100% | 100% | ~99.9% | ~98.4% | ~99.1% |
| Response consistency | Variable (ISP-dependent) | High | High | High | High |
| iMessage | ✓ | ✗ | ✗ | ✗ | ✗ |
| Voice Wake | ✓ | ✗ | ✗ | ✗ | ✗ |
| Data sovereignty | ★★★★★ | ★★★★☆ | ★★★★☆ | ★★★☆☆ | ★★★★☆ |
| Skills compatibility | 100% | 100% | ~92% | 100% | ~90% |
| Best for | Power users | Developers | Cloud-first users | Budget / free tier | Dev + managed |
What We’d Actually Do
If you’ve never run an agent before: start with OpenClaw Cloud. The free tier gives you 14 days to figure out if this is something you actually want, without spending anything or touching a terminal.
Once you know you want it running reliably: TinyClaw is the cleanest upgrade. Isolated containers, consistent performance, fastest onboarding. If you eventually want multiple agents or team access, the infrastructure is already built for that.
If data sovereignty is the point, or you need Voice Wake, or iMessage matters: self-host. The Mac Mini is the obvious hardware choice — $599, silent, 3–4W at idle, runs macOS natively. No cloud option can give you what local-first does.
If you’re a developer who wants to understand the mechanism: read the jlia0 TinyClaw code before you do anything else. 400 lines of shell will tell you more about what’s happening than any documentation.
The Honest Assessment
“Working” and “useful” are not the same thing. Every option we tested was working within its first hour. Useful took longer — it required SOUL.md to accumulate context, it required connecting the tools you actually use, and it required giving the agent tasks that matter rather than demo prompts. The hosted options make it easier to start. The self-hosted options give you more to work with once you’re past the start.
The Skills ecosystem is the biggest open question. Security firm Koi Security audited ClawHub in early February 2026 and found 341 malicious Skills out of 2,857 analyzed — about 12% of the catalog, most of them from a single attacker account that published 314 poisoned Skills in one week before being caught. OpenClaw has since added review processes, but the catalog is too large to audit manually. Treat community Skills the way you’d treat browser extensions: useful, often excellent, and worth a quick look at the source before you install anything that asks for system access.
207K GitHub stars means something went right. Whether this specific thing was right for you is the question that takes a week to answer, not a landing page.
Part 10 of the Agent Infrastructure Series. Previous entries covered architecture, skills ecosystem, enterprise workflows, the ClawdBot deep dive, business models, what it means when its creator joins OpenAI, and deployment guides.
References:
- OpenClaw GitHub Repository (207K stars, v2026.2.17)
- TinyClaw — One-Click OpenClaw Deployment
- OpenClaw Cloud
- EasyClaw.ai
- TinyClaw (jlia0)
- CVE-2026-25253 — OpenClaw Gateway Vulnerability (referenced in our security analysis)