It is eleven at night. You close the laptop with three things still unfinished: the research due tomorrow morning, a price you have been watching for a week, an inbox that keeps growing. The 2026 wave of agent launches is a bet that you will hand these things to something that never clocks out. OpenAI shipped its always-on agent Dots at DevDay 2026. Manus launched Cue, a standalone app. The open-source community pushed feder-cr/dots past 2,400 stars within a day of appearing. And two earlier contenders are still on shift: Devin, already embedded in engineering teams, and OpenClaw 2.0, which takes a local-first route.
This article puts all five at the same table and asks one question: who takes the night shift for you, and at what cost. By "always-on" we mean the agent still holds your task after you walk away: it keeps its context, keeps working while you sleep or sit in meetings, and comes back to you when something needs attention. That is a different trust structure from a one-question chat window: you are delegating a standing responsibility, not a query.
Scope: what this review covers, and what it does not
Two existing reviews on this site share this territory, so boundaries first. ai-computer-use-agents-comparison-review covers computer-use agents in depth; agent-harness-comparison-review covers coding harnesses. This review focuses on resident, personal agents: agents whose value lies not in any single benchmark score but in whether they deserve long-term custody of your tasks. We therefore do not re-run coding benchmarks or run a computer-use deep dive; we compare deployment form, identity and credentials, cost structure, and openness.
Ground rules: costs cannot be compared directly
The five cost structures are completely different, and this is the single most important premise of the whole comparison. Dots is bundled into ChatGPT subscription tiers. Cue is free during an invite-only early period. feder-cr/dots runs on OpenRouter tokens that you pay for directly. OpenClaw reuses subscriptions and API keys you already own. Devin publishes no public price at all. Subscription quotas, invite-based free access, and per-token billing are three currencies that do not convert into each other, so this article describes them in relative terms only. The same discipline applies to capability claims: everything below comes from vendor materials or project READMEs, with no common yardstick, so we describe form factors rather than rank winners.
OpenAI Dots: the first officially hosted night guard
Dots was the headline of OpenAI's new way of working at DevDay 2026 on September 29. Official wording describes it as "a highly capable, always-on agent": it learns what matters to you, works on your behalf continuously, and takes over important work to give you your time back. Its defining trait is hosting: you install nothing, because it lives inside the ChatGPT subscription system. Availability is equally explicit: it targets Pro and Business Premium tiers in eligible markets, while Enterprise, Edu, and Healthcare workspaces can enable it as a beta through their administrators, off by default.
Two calibrations. First, the widely repeated phrase about "long-term task tracking, proactive reporting, and autonomously operating software" is a media paraphrase, not verbatim recap language; treat it with a discount. Second, Dots carries no separate price tag; its cost lives inside the subscription tier. Hold a qualifying tier and it is the lowest-friction of the five; otherwise it is the one that requires buying a ticket first. The hosted form cuts both ways: zero installation and zero maintenance, with upgrades and security handled by OpenAI, at the price of data, credentials, and execution all living in someone else's house. For OpenAI's broader agent-engineering push, including the Codex cloud loop, see openai-codex-agent-work-hotspot.
Manus Cue: an agent with its own identity papers
Cue takes the opposite path: instead of hiding inside a subscription, it ships as a standalone app on mobile and desktop, sharing infrastructure with Manus. Its most radical choice is identity: per the official Manus blog post of September 28, each Cue agent gets its own email address, its own phone number, its own wallet, and its own computer. That means it can pay for things within a budget you set, answer your calls and leave summaries, and even scan a restaurant QR code to order food or hold a place in line.
The independent wallet deserves its own paragraph. A separate wallet turns "how much is it allowed to spend" into an explicit, budgetable parameter, instead of using your card or demanding manual confirmation at every step. Multi-agent group chat makes the division of labor visible: in the official example, three agents split the work of researching venues, filtering a candidate list, and drafting a presentation, and you watch them coordinate instead of staring at a black box.
On cost: Cue is in an early-access period, free with an invite code on a first-come, first-served basis. The official invite code is MEETCUE, and invited users can extend further invitations. Web, desktop, and mobile versions are live, while the iOS app is still pending App Store review. A caution: "free" is the language of an invite-based early period, not a long-term pricing promise.
feder-cr/dots: the open-source understudy
The open-source version appeared one day after the closed-source launch. feder-cr/dots was created on September 29, 2026, and by the October 2 snapshot it had 2,416 stars and 421 forks under an MIT license. The README frames its philosophy plainly: every AI agent is a model plus a browser. You can swap the model with one parameter; the browser is what the website actually sees.
It lives in an environment you host yourself. Models come through OpenRouter, swappable with a single --model flag, and you pay the token costs directly. The README's example task captures its temperament: check the cheapest one-way economy fare between Milan and Lisbon on a travel site, verify each day from the 12th to the 16th of next month, and refuse to invent numbers when no fare exists. That is a textbook night-shift task: tedious, repetitive, and honesty-critical. On the browser side, fingerprints are decided at engine level, --seed reproduces the same identity on every run, and --profile-dir preserves logins and cookies across sessions. One thing must be stated clearly: the README closes with "Not affiliated with OpenAI. MIT licensed." It has no relationship with OpenAI's official Dots; the shared name is a coincidence of naming. A compliance boundary belongs here too: an anti-detection browser agent should only ever automate accounts you are authorized to access and your own business, in line with the target site's terms of service and local law, and never for fraud, scalping, or bulk registration.
OpenClaw 2.0: the local-first gateway route
OpenClaw 2.0 (v2026.8.1) shipped on August 31, 2026, built by 933 contributors across more than 16,000 pull requests; our news piece covers the release in depth. Its architectural keyword is Gateway: session state, credentials, permissions, and persistent work are held in one place, while the browser, the CLI, paired devices, and remote workers all act as entry points. Many doors, one state.
Its posture inverts that of Dots and Cue: rather than building identity from scratch, it starts from assets you already have. Existing ChatGPT or Claude subscriptions, API keys, and local models served by Ollama or LM Studio can all plug in. Two security designs stand out: routes are verified against reality before being saved, and credentials are referenced rather than silently moved. For a broader comparison of how agent products handle credentials and permissions, see agent-credential-permission-comparison-review. One warning is worth repeating: Incognito mode only skips the transcript; it does not stop requests to the provider, so do not mistake it for a privacy mode. On cost: it runs on subscriptions and keys you already pay for, and OpenClaw itself adds no separate bill. Background in openclaw-2-0-release-hotspot.
Devin: the cloud software engineer with the heaviest security stack
Devin is the least "personal assistant" of the five. Its self-assigned job title is the autonomous software engineer, per Cognition's official site: it runs in the cloud or on your machine, tests in its own browser, and keeps going until a pull request is mergeable. It can spin up Windows, Mac, or Linux virtual machines on demand for builds and tests, recording the screen as it clicks. It can mix and match frontier models, with official claims of savings up to 60 percent. It splits large tasks across parallel cloud agents that keep working after you close the lid. Mention @Devin in Slack or Teams and it reads the thread and gets to work. It watches for bugs and broken CI and fixes them proactively; Devin Review hunts for bugs and vulnerabilities before release, and Security Swarm sweeps for security holes, performance issues, and test-coverage gaps.
Its differentiator is security and compliance: input and output IP belongs to the user, zero data retention is enforced toward model providers, audit logs are kept, each session runs in its own VM, RBAC, SSO, SCIM, and OIDC are supported, and private networking is optional. This is the language of enterprise procurement checklists. On price, honesty requires a plain statement: the website snapshot we reviewed publishes no price figure, so this article writes none; contact the vendor for quotes. Devin is also one of the sixteen launch partners for OpenAI's Sign in with ChatGPT credit interoperability (official recap).
One table, five form factors
| Agent | Where it lives | What it does for you | Identity and credentials | How it is paid for | Openness |
|---|---|---|---|---|---|
| OpenAI Dots | Officially hosted inside the ChatGPT subscription system | Learns what matters to you and works on your behalf continuously (official wording) | Rides on the ChatGPT account system | Bundled with subscription tiers (Pro and Business Premium) | Closed source |
| Manus Cue | Standalone app on mobile and desktop | Research, payments within budget, call answering with summaries, multi-agent group chat | Each agent has its own email, phone number, wallet, and computer | Free during invite-only early access (official code MEETCUE) | Closed source |
| feder-cr/dots | Open source, self-hosted in your own environment | Web-task agent with its own browser; model swappable in one flag | Engine-level fingerprinting; --seed keeps one identity; --profile-dir keeps logins | You pay OpenRouter token costs directly | Open source, MIT, not affiliated with OpenAI |
| OpenClaw 2.0 | Open source, self-hosted Gateway, local-first | Centralizes state, credentials, permissions, and persistent work behind many entry points | Credentials referenced, not silently moved; routes verified before saving | Reuses subscriptions and API keys you already have | Open source |
| Devin | Primarily cloud, with local option | Autonomous software engineer; works until the PR merges, watches CI, fixes bugs, finds vulnerabilities | Per-session VM, audit logs, RBAC/SSO/SCIM/OIDC | No public price in the site snapshot; none written here | Closed source |
Read the payment column with care: the five entries cannot be converted into one another. The table's value is matching yourself to a form factor, not ranking.
Picking your night guard, by user type
Already holding ChatGPT Pro or Business Premium and wanting the lowest friction: Dots is the default answer, with nothing to install and no new identity to manage. The trade-offs are closed-source hosting, availability tied to tier and market, and capability claims that partly rest on media paraphrase.
Wanting to outsource errands and try new things: Cue has the most complete independent-identity design, and the separate wallet gives delegated spending a clear authorization boundary. The free invite period lowers the cost of experiment, provided you accept early-product instability and pricing uncertainty.
Developers who like to tinker: feder-cr/dots offers the most freedom, with a swappable model, a self-controlled environment, and a permissive license; the costs are token bills, environment maintenance, and a compliance boundary you police yourself.
Sensitive data with existing subscription assets: OpenClaw's local-first route keeps data and credentials on your side of the fence, and the Gateway architecture suits multi-device, multi-entry-point users.
Engineering teams with procurement requirements: Devin is the only one designed against an enterprise security stack, where zero data retention, isolated VMs, and audit logs are hard requirements. The trade-off is a non-public price and a sales process.
One line to close: these five night guards share no common pricing language. Subscription quota, invite-based free access, and per-token billing remain three currencies. The real question of this comparison is not which agent is strongest, but what you are willing to hand over, and what you insist on keeping for yourself. Five agents on the night shift: which one gets your three-in-the-morning task?
FAQ
Q1: Can the five agents' costs be compared side by side?
No. Subscription tiers (Dots), invite-based free access (Cue), self-paid tokens (feder-cr/dots), reused subscriptions (OpenClaw), and unpublished pricing (Devin) are different structures that cannot be converted into a single number. This article describes them only in relative terms.
Q2: How do I get access to OpenAI Dots right now?
Per official wording: it targets Pro and Business Premium tiers in eligible markets, while Enterprise, Edu, and Healthcare workspaces can have administrators enable the beta, which is off by default.
Q3: Does Cue cost money?
Per the official Manus blog: Cue is in an early-access period, free with an invite code on a first-come, first-served basis. The official code is MEETCUE, and invited users can invite others. The iOS app is pending App Store review. This is an early-period arrangement, not a long-term pricing commitment.
Q4: Is the open-source dots the same thing as OpenAI's Dots?
No. The README of feder-cr/dots ends with "Not affiliated with OpenAI. MIT licensed." It has no relationship with OpenAI's official product; the names merely match, and the license is MIT.
Q5: How much does Devin cost?
The Cognition site snapshot we reviewed contains no public price figure. Following this article's sourcing discipline, no price is written here; request a quote through official channels before purchasing.