OpenAI Dots vs OpenClaw: Which Personal Agent Actually Fits in 2026?

OpenAI's dots and the open-source OpenClaw both want to be the agent that keeps working after you close the laptop. One is a polished add-on to a ChatGPT plan; the other is a lobster you host on your own hardware. This walkthrough lines them up on the same axes — where they run, what they cost, what they can and can't do — using footage from three hands-on reviewers: Theoretically Media's DevDay coverage, Mark Kashef's dots deep-dive, and Rachel How's two weeks self-hosting OpenClaw. Every verdict below belongs to the reviewer who said it; the final call is yours.

Source & credits

Screenshots in this guide are captured from Theoretically Media's public walkthrough video. Every step links back to the exact moment it shows, so you can follow along.

Theoretically Media ↗

Two answers to the same job

  1. 1

    Meet the two contenders

    OpenAI used DevDay 2026 to introduce dots: persistent agents that keep working whether or not ChatGPT is open. The timing matters. Reviewers immediately framed dots as a hosted answer to OpenClaw, the open-source agent whose creator, Peter Steinberger, joined OpenAI in February. The next slides show why the comparison stuck.

    OpenAI DevDay 2026 title card where the dots persistent agent was announced to a live audience
    DevDay 2026 is where dots went public.Watch at 0:25
  2. 2

    What dots promises

    Each dot runs 24/7 with its own cloud computer, and you can reach it by chat, voice call, the mobile app, or Slack and Teams — and it can reach back. Theoretically Media's read: these are "bots with timers and tasks", not a conscious co-worker. It is the same always-on promise OpenClaw users already run on their own machines.

    dots wordmark on the DevDay stage screen beside a phone demo of chatting with your dot
    The pitch: an agent you can call, text, or ping on Slack.Watch at 1:15
  3. 3

    The OpenClaw side of the story

    OpenClaw didn't stand still while OpenAI built its take. Its creator joined OpenAI, and the project itself became a foundation — open and independent, per Steinberger's own post, and "just getting started". That backdrop is why the two keep getting compared: dots is, structurally, a managed version of the thing OpenClaw lets you run yourself.

    X post by Peter Steinberger announcing he joined OpenAI as OpenClaw becomes an open foundation
    February 2026: Steinberger heads to OpenAI; OpenClaw becomes a foundation.Watch at 1:40
  4. 4

    First-run: dots inside the ChatGPT app

    dots lives in the ChatGPT/Codex app, and onboarding asks you to connect Email, Calendar, Contacts, and Drive. Mark Kashef notes that integrations you already set up in Codex don't carry over — you reconnect them here. OpenClaw's first run instead interviews you about personality, tone, and memory structure. The two philosophies diverge from minute one.

    Codex app Make your dot more helpful screen offering Email Calendar Contacts and Drive connectors
    dots onboarding: reconnect your Google stack inside Codex.Watch at 0:40

Where the agent actually runs

  1. 5

    dots: a computer on OpenAI's servers

    Every dot gets an isolated cloud computer on OpenAI's infrastructure, driven by hooks for event triggers and cron for scheduled jobs. You never see the machine — the diagram below is Theoretically Media's simplification — you only see chat and results. OpenClaw inverts this: the machine is the Mac mini or VPS humming in your own home or account.

    Simplified illustration of a dot cloud computer on OpenAI servers with hooks cron and AI model tools
    Hooks and cron on OpenAI hardware, isolated per dot.Watch at 5:18
  2. 6

    Peek inside a dot's computer

    The profile panel shows what a hosted agent looks like in practice: groknot is active, has a Call button and a Slack handle, and — the key line — its own computer, marked Connected. When the dot drives that computer mid-task you can watch it work or press Take over. With OpenClaw you would simply open a terminal on the box itself.

    dots profile panel showing groknot active with a Call button and groknot's computer Connected
    groknot's computer reads Connected — that's the whole remote VM.Watch at 1:05
  3. 7

    OpenClaw: your hardware, reachable anywhere

    Rachel How runs OpenClaw on a used Mac mini that stays on 24/7 and reaches it from her phone or laptop through Tailscale, whose free plan is enough. Remote access to your own agent costs nothing extra because the agent was never anywhere else. The trade: you are the datacenter now, from electricity to uptime.

    Tailscale admin console listing machines that reach a self-hosted OpenClaw Mac mini from anywhere
    Tailscale's free plan bridges her phone to the Mac mini.Watch at 3:17

What you actually pay

  1. 8

    dots pricing: 100, 200, Pro 500

    DevDay reshuffled ChatGPT Pro into $100, $200, and a new Pro 500 tier with 25× Plus usage and Ultrafast access. dots requires one of these tiers — the $20 Plus plan doesn't include it. Theoretically Media, himself on Pro, still hadn't received his dot at review time, so treat availability as rolling even on qualifying plans.

    DevDay slide comparing ChatGPT Pro tiers at 100 200 and Pro 500 with 25x Plus usage for dots
    Pro 500 tops the new lineup at 25× Plus usage.Watch at 7:50
  2. 9

    The plan gate: Pro, Business, Enterprise

    The "dots in Space" slide doubles as the eligibility list: ChatGPT Pro, Business Premium, or Enterprise. That gate is the real price of the polished experience — you're not buying dots, you're upgrading your whole ChatGPT plan to reach it. OpenClaw's gate is a different toll: whatever hardware and API key you bring.

    dots in Space DevDay slide confirming dots requires ChatGPT Pro Business Premium or Enterprise
    Check the checkmarks before you budget for dots.Watch at 6:35
  3. 10

    OpenClaw costs: metered, itemized, yours

    Self-hosting isn't free — it's itemized. Rachel How's stack: a used Mac mini, about a dollar of monthly electricity, ChatGPT Plus via OAuth or an OpenRouter key, and a few add-on APIs like Brave Search at roughly $5 per 1,000 queries. Nobody charges a platform fee; you pay for the parts you use, and you can unplug any of them.

    Brave Search API pricing page showing 5 dollars per 1,000 queries for the data for AI tier
    Brave Search: about $5 per 1,000 queries, one line in the stack.Watch at 3:38

Where capabilities split

  1. 11

    dots at work: build tasks with guardrails

    Ask dots what it's good at and it lists research, documents, slides, websites, code, and travel comparisons — and promises to ask before purchases or sharing sensitive information. Kashef's build-me-a-website task came back with a plan, a green checkmark, and a "link when it's ready" message. The polish is real; so is his complaint that it waits to be pushed.

    dots capability list promising to ask before purchases beside a build chatgpt site task checkmark
    A green checkmark on a build-me-a-site task.Watch at 3:05
  2. 12

    Computer use, with a Take over button

    Given a Montreal-to-Johannesburg flight search, groknot browses the Emirates site inside its own computer window — visibly, slowly, under a "groknot has control" banner with a Take over button. Kashef's verdict on the demo: competent but slower than running the same task in Codex, and you watch it happen rather than receive a done result.

    dot named groknot has control of its cloud computer browsing Emirates for a Montreal Johannesburg flight
    dots drives the browser; you can grab the wheel.Watch at 3:48
  3. 13

    dots in Slack: three clicks, its own profile

    Connecting dots to Slack takes about three clicks, after which it appears as an agent with its own profile and threads. In Kashef's test it answered "how many emails did I get in the past hour?" straight from connected Gmail. It replies in threads and can be @-mentioned in channels — the coworker simulation is deliberate.

    mark-groknot Slack agent profile answering how many emails arrived in the past hour from Gmail
    mark-groknot answers an email-count question in a Slack thread.Watch at 4:40
  4. 14

    OpenClaw at work: tools you build yourself

    Rachel How's first big OpenClaw project was a mission control dashboard — calendar, per-agent context usage, and task tracking — which she says the bot built for her in about five minutes. Morning rundowns arrive by message, and YouTube analytics pull in through the API. Nothing about it is polished until you or your agent make it polished.

    OpenClaw mission control dashboard with Tasks columns tracking context usage and agent to-dos
    A mission control kanban the bot assembled in minutes.Watch at 1:35

Which one fits you

  1. 15

    The limit dots admits itself

    Asked whether you can spin up a team of dots, the product answers in its own words: it can coordinate tasks in parallel, but "I can't create separate dots from this chat." One dot per account for now, with subagents working inside a chat. OpenClaw has no such ceiling — as many agents as your hardware and patience allow. More dots are promised later; the ceiling is real today.

    dots chat replying that it cannot create separate dots when asked to spin up a team of dots
    dots itself says: no separate dots from this chat.Watch at 1:35
  2. 16

    The filter, not a verdict

    Rachel How's table is the fairest closing frame: you'll love OpenClaw if you enjoy tinkering, Google things when they break, and are happy to spend a week setting up — and it's not for you yet if you want things to just work or need results immediately. Reviewers' own scores: ZDNET called dots "OpenClaw declawed — for $100 a month"; Kashef says don't upgrade plans for one feature; Theoretically Media expects this battleground to stay interesting. Pick your row.

    OpenClaw fit table comparing You'll love it versus Not for you yet across attitude breakage and time
    Rachel's fit table — the decision we leave with you.Watch at 0:37

Frequently asked questions

Keep exploring