Agent comparisons

OpenAI Dots vs OpenClaw: Always-On Agents, Controls, and Cost

Choose OpenAI Dots if you are on ChatGPT Pro or Business Premium, work in Slack or Microsoft Teams, and want OpenAI to run the agent with its own computer, thousands of app connectors, and approval rules. Choose OpenClaw if you want to pick the model, edit the files that drive the agent, decide how often it wakes up, run several agents, and reach it on WhatsApp, Telegram, or Slack. Both are agents that keep working after you close the chat. The difference is who holds the controls. Lobsterland runs OpenClaw as a managed instance, so the open option does not mean running a server.

Deploy OpenClaw See hosting costs

By Lobsterland · Reviewed October 16, 2026

We host OpenClaw and Hermes. Dots launched on September 29, 2026. The Dots facts below were checked against OpenAI’s announcement and help documentation on October 16, 2026. We have not tested a Dots account; account eligibility and controls depend on the rollout and workspace settings.

First, the names. Dots is OpenAI's always-on personal agent, available through ChatGPT and powered by GPT-6 Astra. It is not the Agents SDK for developers. OpenClaw is the open-source agent runtime. It is not OpenClaw Launch or any other hosting company that uses the name.

Is OpenAI Dots built on OpenClaw?

OpenAI’s Dots announcement describes an OpenAI product, not an OpenClaw distribution. Similar always-on behavior does not establish shared code. OpenClaw remains a separate open-source runtime.

The main difference is openness. Dots uses OpenAI’s model and managed cloud computer, with optional access to a computer you connect. OpenAI sets the product’s rules. OpenClaw leaves those parts open: which model answers, which instruction files it reads, how often it wakes up, and which channels it listens on. That openness is useful when you want to change something. It is work when you would rather not decide.

OpenAI Dots vs OpenClaw: what actually differs?

The Dots column follows OpenAI’s announcement and current Dots help documentation. The OpenClaw column describes managed OpenClaw hosting on Lobsterland.

How does each one work without being asked?

Dots can pursue background goals and perform read-only proactive research. You can also ask it to schedule a recurring check or reminder and change its repeat timing in Scheduled. That is distinct from the background research the product initiates itself. See OpenAI’s scheduling instructions.

OpenClaw gives you a configurable heartbeat monitor and separately scheduled automation jobs. On the 2026.10.2 image used by Lobsterland, the monitor checklist lives in monitor scratch, rather than HEARTBEAT.md. Scheduled heartbeats require automations to be enabled. For a task that must run at a particular time, create an automation with an explicit schedule. Our OpenClaw cron guide explains the scheduling use cases; the versioned heartbeat documentation covers the current monitor behavior.

For a standing goal, tell the agent what to check, where the evidence is, and when to notify you. For example: check the monitoring feed for a new error class, report the evidence and one next step, otherwise stay quiet. Have it create a scheduled automation for a recurring task and inspect the saved schedule. A goal written in a workspace file alone does not establish a schedule.

Lobsterland configures a 30-minute heartbeat interval by default: 48 potential intervals per day, not 48 guaranteed model calls. Polls can be skipped without calling the model; actual model work consumes your provider allowance or metered usage. Set a longer interval such as 2h to reduce frequency. 0m disables recurring heartbeats, but event-driven completion turns can still run. Our generated configuration uses isolated monitor sessions and target: last for heartbeat alerts. Completion events follow the session that owns the work, so these settings are not a promise that every background result goes to your most recent chat.

One dot, or several agents?

OpenAI includes a first dot with Pro and Business Premium. Its announcement presents teams of dots as a future direction and specialist dots with organizational identities and defined responsibilities as a preview. Treat those separately from the primary dot you can use today.

OpenClaw can do the routing half of that today. One instance can run several agents. Each has its own workspace files, name, and model. A binding sends a specific conversation to a specific agent: one person's Telegram or Slack direct messages, a Telegram group, or a Slack channel. A small team could run a support agent in the #support channel, a finance agent in #invoices, and a personal agent for each person's DMs, all on one instance. Multi-agent routing is set up in the dashboard, without editing config files.

The trade-off is that you design the split and write each agent's files. OpenClaw bindings decide which agent answers a conversation. They are not a manager that hands work between agents on its own.

Where do you talk to it?

Dots supports ChatGPT, Slack and Microsoft Teams. You can call your dot; it cannot initiate calls at launch. Texting is a limited beta for some US Pro users, unavailable in Business and Enterprise workspaces. Check the current channel instructions and your account before planning around it.

Managed OpenClaw offers built-in chat in the dashboard, the official OpenClaw mobile app, WhatsApp, Telegram, and Slack. WhatsApp is what Dots does not have. You can talk to the agent in your own self-chat, or give it a dedicated number of its own, and a DM policy plus an allowlist decides who else can reach it. The WhatsApp setup guide covers each step.

Lobsterland has no Microsoft Teams connection, and no phone or SMS. If your company works in Teams, that one fact may decide the question for you.

What stops it before a risky action?

Dots has Custom Rules and auto-review for actions. Built-in safeguards still apply, and some steps stay with you. Enterprise administrators control whether members can use Custom Rules; app permissions remain separate. The workspace documentation explains those boundaries. These are documented controls, not a guarantee that an agent cannot make a mistake.

Managed OpenClaw works differently. Elevated tools, the ones that run commands, run in full mode by default. They can be triggered by the people your channel settings let in, and they do not pause for a confirmation on each action. So the main control is who can instruct the agent, not an approval screen for each step. That makes "allow anyone" a serious setting. If you open a WhatsApp number to everyone, everyone who messages it can trigger those tools. You can write "draft, do not send" into AGENTS.md, but model compliance is not guaranteed. That is an instruction, not an enforced rule.

In practice: keep allowlists tight, have the agent report in a channel you read, and don't sign the Hosted Browser into accounts where a mistake would be costly. If you need configurable action approvals, Dots offers more built-in controls; verify the relevant rule and app permissions before relying on them. Our security overview explains where the managed boundary stops.

Where is Dots ahead of OpenClaw?

Approval controls. Per-action rules, an activity view, and an auto-review are built in. On OpenClaw, you would have to design those yourself.

Its own computer and connectors. A dot has a cloud computer and browser of its own, plus more than 4,000 apps through OpenAI's plugins. OpenClaw works through a Hosted Browser you sign into and the skills you install.

Teams and integrated setup. Dots supports Microsoft Teams and uses an eligible ChatGPT plan. You still connect apps and review permissions, but you do not choose a model or administer a runtime.

Enterprise identity. Specialist dots get their own organizational identity and credentials, and OpenAI engineers are working with pilot customers to set them up. OpenClaw has no equivalent managed offering.

What do access and cost look like?

As checked on October 16, 2026, the first dot is included in eligible Pro and Business Premium plans, with a deeper-work allowance and extended first-month limits. Pro excludes the EEA, Switzerland and the UK; Business Premium covers supported regions. Enterprise, Edu and Healthcare require admin-enabled beta access. Consult OpenAI’s current eligibility and allowance documentation for your account.

Managed OpenClaw on Lobsterland starts at $6.90 per month. Model usage is billed by your connected provider, or uses your ChatGPT/Codex allowance with supported login. For Claude, use an Anthropic API key as the default: hosted subscription authentication may incur additional metered usage and should not be treated as covered by a Pro or Max plan. Using your ChatGPT subscription with hosted OpenClaw explains how that login works. It powers OpenClaw with OpenAI models, and it does not bring your dot along. Your heartbeat interval is one of the settings that changes your model bill. Every plan starts with a 7-day trial, and a card is required up front. The pricing page shows hosting and model access separately, so you can estimate both. If you already pay for Pro, Dots may cost you nothing extra. OpenClaw would then be an added cost you pay for control.

How do you try OpenClaw without a server?

1. Create a Lobsterland account and choose OpenClaw for a new instance.
2. Connect a model API key or supported ChatGPT/Codex login. Use an Anthropic API key for Claude.
3. Write USER.md (who you are, how you like answers), then ask the agent to create one recurring automation with a clear source, schedule and notification rule. Inspect the saved job before relying on it. Our guide to OpenClaw's memory files explains each file.
4. Connect Slack, Telegram, or WhatsApp, set the heartbeat to an interval you can live with, such as 2h, and watch a day of results before adding more jobs or a second agent.

A fair first test is the one OpenAI leads with: an alert to investigate. Give the automation access to the feed where alerts land. Ask it to say what changed, quote the evidence, and propose one next step without acting on it. If it misses an alert or repeats itself, change the checklist and run it again. Being able to make that edit is the main thing OpenClaw gives you.

If you only need chat, Telegram or Slack, and a runtime with readable skills, OpenAI Dots vs Hermes Agent covers that narrower choice. For Meta's take on the same idea, see Meta Muse vs OpenClaw. If you are weighing plain ChatGPT rather than Dots, read ChatGPT vs OpenClaw.

Frequently asked questions

Is OpenAI Dots built on OpenClaw?

OpenAI describes Dots as its own product, not as an OpenClaw distribution. Similar always-on behavior does not establish shared code. OpenClaw is an open-source runtime you can run yourself or on a managed host.

Can I use my ChatGPT subscription with OpenClaw?

Yes, as the model. Managed OpenClaw supports ChatGPT/Codex login, subject to your plan's limits. That gives OpenClaw access to OpenAI models. It does not bring your dot, its cloud computer, or its app connectors.

Can OpenClaw run several agents like a team of Dots?

One managed instance can run several agents today, each with its own workspace files, name, and model. Bindings send a person's Telegram or Slack messages, or a group or channel, to a chosen agent. OpenAI describes teams of Dots as a future direction.

Does OpenClaw work in Microsoft Teams or over the phone?

Not on Lobsterland. Managed OpenClaw connects to built-in chat, WhatsApp, Telegram, Slack, and the OpenClaw mobile app. There is no Teams connection, phone calling, or SMS.