TL;DR

  • Every big vendor now ships an agent that works across your apps: Claude, ChatGPT Work, Gemini Spark, Siri in iOS 27, Microsoft’s Copilot Cowork. The open-source side has OpenClaw and Hermes Agent.
  • They differ less in intelligence than in where they run, where your data lives, and how much damage one bad action can do.
  • I ran OpenClaw in my cluster from 2026-02-22 to 2026-09-13, then archived it and deleted the gateway on 2026-10-20. Meanwhile 1,526 of 1,624 commits this year in my busiest repo were co-authored by Claude.
  • There probably isn’t a best assistant. There’s the one your own usage data says you actually use. If that data points at something nobody sells, build it. If it doesn’t, don’t.

Why “which one is best” is the wrong question

Every comparison post I read ranks these tools on a feature grid. Memory: yes. Computer use: yes. WhatsApp: some. The grid makes them look interchangeable, and it hides the three things that actually decided whether I kept using one:

  1. Where it runs. On your laptop (stops when the lid closes), on a server you own (you are now the on-call), or on the vendor’s cloud (always on, and your data lives there).
  2. What it can touch. An assistant with your inbox, your files and a shell is useful exactly because it’s dangerous. Palo Alto Networks calls private data plus untrusted content plus the ability to send things out a “lethal trifecta”. Every option below has all three.
  3. Whether it finishes the work. The Remote Labor Index gives agents 240 real freelance projects worth $144,000. When it launched, the best agent completed 2.5 percent at a quality a client would accept. By July 2026 the best score was 16.1 percent. That’s a huge jump, and it still means most real jobs don’t come back done.

So the honest framing is: pick the failure modes you can live with.

The list

Pricing moves monthly and several vendor pages wouldn’t load for me, so I’ve only listed prices I could see on the vendor’s own page. Check before you buy.

OpenClaw (open source, self-hosted)

The one everybody’s heard of. Formerly Clawdbot, then Moltbot. It runs on your own machine and talks to you through the chat apps you already use: Discord, iMessage, Slack, Teams, Telegram, WhatsApp, Signal and more. MIT licensed, any model you want. In July it became a nonprofit foundation. Its creator joined OpenAI earlier in the year and, per the announcement, still makes the technical calls.

Pros

  • Widest channel coverage of anything here. It meets you where you already type.
  • Any model, local or cloud. Your config, your logs, your box.
  • Free, and you can read every line.

Cons

  • The security record. In February, an audit of all 2,857 skills on ClawHub found 341 malicious ones, 335 of them installing a macOS credential stealer through fake “prerequisites.” The only bar to publish a skill was a GitHub account a week old.
  • Tools run on the host by default unless you set up sandboxing. I spent a whole post hardening my deployment and it was still a lot of attack surface for a chat bot.
  • Analysts quoted by InfoWorld question how neutral a foundation is when OpenAI is described as a major donor. That may never matter to you. It’s worth knowing.

Hermes Agent (Nous Research, open source)

The other self-hosted contender. MIT licensed, runs on macOS, Windows and Linux, remembers projects, writes its own reusable skills, and reaches you over Telegram, Slack and WhatsApp. Nous raised $90M for a business version.

Pros: Same self-hosted, bring-your-own-model appeal as OpenClaw, with self-written skills as the headline feature. Cons: I couldn’t find independent security reviews either way. “No incidents reported” and “nobody has looked yet” read the same from the outside. And self-written skills are code your assistant wrote that runs later with your permissions, so you’ll want to review them.

Claude (apps, Cowork, Claude Code, Agent SDK)

What I actually use every day. The pricing page lists memory and connectors on every plan including Free, Claude Code on Pro and up, and says Cowork, the desktop agent that works in your files, is now just part of Claude. Claude Code is the terminal agent. The Agent SDK is the same harness as a library, for building your own.

Pros

  • The best fit I’ve found for work that lives in files and repos: code, infrastructure, documents.
  • Extensible in boring, inspectable ways: skills are markdown files, tools come in through MCP servers, behaviour is set in a CLAUDE.md. I have 46 skills on my workstation.
  • The SDK means “build your own” doesn’t start from zero.

Cons

  • The desktop agent runs on your desktop. When the laptop sleeps, the assistant sleeps.
  • No native WhatsApp or Telegram presence that I could find. I bridged it to Mattermost myself.
  • Anything that can edit your files on instruction can edit them on a bad instruction. Coding agents had a rough summer for prompt injection (more in the security roundup).

ChatGPT Work (OpenAI)

OpenAI retired the original ChatGPT agent and replaced it with ChatGPT Work in July. It covers connected apps, files, scheduled tasks, a desktop browser and computer use, and folds the Codex app into the ChatGPT desktop app.

Pros: The biggest consumer install base, so the smoothest on-ramp for non-technical people. Scheduled tasks are built in. Cons: Cloud-hosted means your connected data lives in OpenAI’s cloud. OpenAI’s own pages blocked my fetches, so I can’t quote plan limits.

Gemini Spark (Google)

Announced at I/O in May: a 24/7 personal agent that runs on dedicated cloud VMs, so it keeps working when your laptop is closed. It works across Workspace and asks before high-stakes actions like spending money or sending email. Google also shut down Project Mariner, its earlier browser agent, and folded it in.

Pros: Always-on without running a server yourself. If your life is already Gmail, Calendar and Docs, the integration is the whole point. Cons: It started as a limited, US-only rollout tied to the top-tier subscription. It runs on Google’s VMs, with Google’s data.

Siri in iOS 27 (Apple)

The new Siri holds a conversation, has its own app and history, and can search and act across Mail, Messages, Notes, Reminders and Calendar. Apple’s next-generation models are reported to be based on Gemini under a multi-year deal.

Pros: It’s already on the phone, it’s free, and it reaches into first-party apps nothing else can touch. Cons: Limited to Apple’s apps and recent hardware. Early reviews called it practical but plain. I couldn’t find a clear public answer on which requests stay on Apple’s servers and which reach Google’s models.

Copilot Cowork (Microsoft)

Generally available since June 16. It needs a Microsoft 365 Copilot licence, bills per task in credits, and is off by default with admin spending caps. It runs on Anthropic and OpenAI models under the hood.

Pros: The right answer if your company already lives in 365 and IT needs controls and billing caps. Cons: An enterprise product. Nobody is buying this for their personal life.

The DIY layer: Home Assistant, Open WebUI, Letta

  • Home Assistant Assist can use any integrated LLM, including a local one, and still handles a fixed command set without one. It’s the best private voice assistant for the house, and it stays narrow on purpose.
  • Open WebUI is a self-hosted chat front end with RAG and MCP. It’s what OpenClaw replaced in my cluster. It’s a good interface, but on its own it’s a chat window, not an assistant that acts.
  • Letta (formerly MemGPT) is a memory-first agent framework if you want to build the assistant yourself and memory is the part you care about.

What my own data said

This is the part a feature grid can’t tell you.

I replaced Open WebUI with OpenClaw on 2026-02-22. I hardened it, wired Claude into it so long-running agents could interrupt me in Slack, and hung dream workers, trading signals and a LiteLLM router off it. On paper it was exactly the assistant I wanted: always on, in my chat app, on my hardware, any model.

On 2026-09-12 I wrote the archive commit: 17 replicas in the openclaw namespace to zero, deploys gated off. On 2026-10-20 I deleted the gateway deployment and its unused personal database outright. The archive commit doesn’t give a reason, and I’m not going to make one up now. What I can show is where the work went instead:

SignalNumberWhere it comes from
Commits co-authored by Claude, busiest repo, 20261,526 of 1,624git log trailers, non-merge commits since Jan 1
Skills on my workstation46ls ~/.claude/skills
Months OpenClaw ran before archiveabout 6.5Feb 22 to Sep 13

The assistant I use is the one sitting in a terminal next to the repos, with skills I wrote for my own workflows and a memory file I keep trimming. Nobody sells that exact thing. I assembled it from parts, on top of a vendor harness, because my usage said that’s where the work happens.

That’s the whole argument of this post. I didn’t build my own because building is better. OpenClaw was a build too, and it lost. The usage data picked the winner, not the architecture.

How to choose, if you’re starting today

  1. Write down three jobs you’d hand off this week. Real ones, like “triage my inbox every morning” or “fix the flaky test.” If you can’t name three, you don’t need an agent yet. A chat window will do.
  2. Match where the job lives. Gmail and Docs point to Gemini. 365 points to Copilot Cowork. Files and repos point to Claude or ChatGPT Work. “Must answer in WhatsApp while I’m out” points to OpenClaw or Hermes, and that means a server you’ll maintain.
  3. Decide what it may not do. Spend money, send email, delete things. The good products already ask before high-stakes actions. Your self-hosted one does whatever you configured, including nothing.
  4. Run it for a month and count. Commits, tasks closed, messages it handled that you’d otherwise have done. If the count is low, switch or stop. That number is the only benchmark that’s about you.
  5. Build only when the count says the gap is real. If you keep doing the same glue by hand and no product covers it, that’s when a skill, a bridge or an SDK agent pays for itself.

Lessons

  • A feature grid ranks capabilities. Your usage log ranks what you’ll actually keep using. Only the second one predicts anything.
  • Self-hosting moves the risk onto you, it doesn’t remove it. The ClawHub incident was a supply-chain attack on people who had chosen to self-host for safety.
  • “Always on” is a real feature, and an expensive one when it means you run the server. Ask whether your jobs actually happen while you’re away before paying for it in money or maintenance.
  • Keep the exit cheap. My OpenClaw archive was reversible by design: PVCs, secrets and images kept, deploy jobs gated. Being able to drop a tool without losing anything made it easy to admit it had lost.