Product comparison · Latest source check
Personal AI agents differ in who runs the computer, chooses the model, and pays for usage. OpenClaw and self-hosted Hermes give you provider control. Hark Pro, Muse, Dots, Grok Bot, Claude, and Manus manage more of the runtime, with different allowances and access rules. Compare those choices before granting access to your files or accounts.
Hark Pro and Muse were checked October 6, 2026; the other six products retain their September 30 checks. We have not installed these products, checked out paid plans, or run matched tasks. Each cell records its own documentation date and unresolved fields.
Scroll horizontally to compare all 8 products. Keyboard: focus the table region, then use the left and right arrow keys. Each cell shows its source-check date. BYOK means connecting your own model-provider API key.
| Decision | Claude | Dots | Grok Bot | Hark Pro | Hermes Agent | Manus | Muse | OpenClaw |
|---|---|---|---|---|---|---|---|---|
| Hosting | Documented Cloud tasks; local files and apps need Claude Desktop open. | Documented Own cloud computer; local computer access is optional and starts off. Dots access and controls Checked | Documented Cursor-hosted cloud computer. Optional local execution has separate controls. | Documented Managed cloud service, including Handoff browser sessions. Web, iOS, and Android interfaces. Hark Pro launchUsing Hark Checked | Documented Self-hosted runtime, or optional Hermes Cloud instance. | Documented Managed agent service; authorized local computer use and a paid Cloud Computer add-on are distinct options. Manus 2.0 and computer use Checked | Documented Dedicated Secure VM cloud computer and browser; web, iOS, Android, and WhatsApp interfaces. Confidential VM is planned. | Documented Gateway on your own infrastructure: local machine or hosted server. |
| Model choice | Documented A model picker remains in the unified experience. | Unknown GPT-6 Astra is documented. A user model picker is undocumented. OpenAI Dots launch and usage Checked | Documented Cursor selects models and failovers. No customer model picker or fixed vendor set. | Unknown Hark uses its own and selected external models. A customer model picker is undocumented. Hark credential security Checked | Documented Configure providers and models, including compatible Ollama and vLLM endpoints. | Unknown Current foundation-model picker and bundled model roster are unverified. Manus 2.0 and computer use Checked | Unknown Muse Spark is documented. A customer model picker is undocumented. Muse launch Checked | Documented Choose a provider/model; custom and local-runtime routes are documented. |
| BYOK and authentication | Unknown Claude account. General third-party model BYOK is undocumented; own-cloud marketing does not establish it. Claude Cowork product Checked | Unknown Eligible ChatGPT account. User-supplied model API keys are undocumented. Dots access and controls Checked | Unknown Cursor account; eligible SuperGrok or X subscription links grant usage. Model BYOK is undocumented. | Unknown Hark account and connected services. Customer-supplied model API keys are undocumented. | Documented Provider keys, documented subscription OAuth routes, or compatible local endpoints. Each route has its own eligibility. | Unknown Manus account. User-supplied model API keys are undocumented in the checked sources. | Unknown Meta account and connected services. Customer-supplied model API keys are undocumented. | Documented Provider API keys and documented subscription routes; local endpoints depend on server authentication. |
| Plan fee and included usage | Documented US Pro: USD 20 per individual subscription/month. Task work consumes plan usage; regional prices differ. | Documented Pro 100/200/500: USD 100/200/500 per individual subscription/month. Business Premium: USD 125 per seat/month, or USD 100 per seat/month billed annually; at least two total Business seats. First dot included; deeper-work allowance applies. | Unknown Eligible Cursor plans or linked subscriptions include usage. Extra usage can be on-demand; exact numeric allowance is unknown. Grok Bot plans and billing Checked | Unknown Free with all features; paid tiers offer 2x or 10x usage. Numeric prices withheld here because displayed dollars do not establish currency; checkout determines it. | Documented MIT software has no license fee. Optional Portal plans include credits; public currency code is unverified, so numeric prices are withheld. | Unknown Credit plans. Current plan amounts and selected allowances are unverified; review the billing interval and credit quantity before checkout. | Unknown Capped free access; Power includes 500 million Muse tokens/week and Maximum 3 billion/week. Numeric prices withheld here because checkout currency is unverified. Account and region offers vary. Muse subscription plans Checked | Unknown No first-party paid tier established in these sources. Provider subscriptions or usage charges are separate. |
| Inference charges | Documented Tasks use subscription allowances. Generic API and Console use is billed separately. | Documented Dot conversations do not consume ChatGPT limits; delegated Codex and ChatGPT Work tasks consume their usual usage. OpenAI Dots launch and usage Checked | Unknown Serving model affects billing. On-demand usage is billed through Cursor; Bot token rates are unverified. | Unknown Product usage allowances apply. Absolute allowance, reset period, and standalone Handoff API rates are unverified. Hark usage plansUsing Hark Checked | Documented BYOK bills the provider. Native Anthropic OAuth uses Max extra credits, not base allowance; Portal uses its credit balance. | Unknown A separate foundation-model inference bill is not established for the consumer product. Task-credit cost per job is untested. | Unknown Weekly Muse-token allowances are product units, not an established API-token or completed-task equivalent. Exact free allowance is unverified. Muse subscription plans Checked | Documented Configured provider and authentication route determine usage charges. Local cost estimates are not provider invoices. OpenClaw usage and costs Checked |
| Tool charges | Unknown Task work uses plan limits. A complete third-party tool bill is unknown. | Documented Delegated work uses its normal allowance. Texting may incur message or data charges. | Unknown A complete separate tool-charge schedule is not established. Grok Bot plans and billing Checked | Unknown Automatic usage top-ups are optional, with a selected ceiling. Exact top-up rates and separate third-party tool charges are unverified. Using HarkHark usage plans Checked | Documented Own tool-provider charges apply by configuration. Nous hosted tools use credits, separate from Cloud hosting. | Unknown A complete separate tool-charge schedule is not established. | Unknown Connectors include Shopify, QuickBooks, Slack, Stripe, and Notion. A complete separate tool-charge schedule is unverified. | Documented Optional search, browser, and other service calls can add their own usage charges. OpenClaw usage and costs Checked |
| Host and hardware charges | Unknown Local tasks require your device. No separate cloud-hosting charge is established here. | Unknown Cloud computer is part of the product; local access uses your device. Future extra-dot prices are unknown. | Unknown Hosted computer is the supported deployment. A separate host charge is not established. | Unknown No separate cloud-hosting charge is established. Planned hardware is not required or a shipped product in these sources. Using HarkHark usage plans Checked | Documented Own hardware or host bill. Hermes Cloud charges uptime plus storage; stopped instances still incur storage. Inference and tools are additional. Hermes Cloud billing Checked | Unknown Cloud Computer is a paid add-on. Its current amount is unknown. Local computer use needs your device. Manus 2.0 and computer use Checked | Unknown The hosted computer is part of the product. No separate hosting charge is established here. | Documented You pay for the chosen server or hardware, storage, power, and upkeep. No common monthly total. |
| Permissions | Documented Connected apps and selected folders; Manual is the default action setting, with Auto optional. | Documented App permissions and Custom Rules control actions. Disconnecting an app does not erase previously collected information. Dots access and controls Checked | Documented One user’s Bots share computer access. Local commands ask for approval by default; configurable policies can allow more. | Documented Credential vault and tokenization. Default confirmation before transactions can be overridden by standing permissions. Provider data use remains a separate question. | Documented Configured user allowlists, dangerous-command approval, and optional file/container restrictions. Their behavior depends on settings. | Documented Local computer use requires a connected, authorized session for approved files, browser, and apps. Manus 2.0 and computer use Checked | Documented Sentinel mediates network and connector access with scoped or standing grants. Purchases require approval each time. Training is enabled by default, with an opt-out. Muse security and data use Checked | Documented Tool policy, approvals, and sandbox settings control access. A shared Gateway is one trust domain. |
| Account and region access | Documented Unified experience rolls out to Pro and Max on web, desktop, and mobile. Account timing varies. | Documented Gradual rollout: Pro outside EEA, Switzerland, and UK; Business Premium in supported ChatGPT regions; Enterprise beta needs admin enablement. Create on desktop. Dots access and controls Checked | Unknown Paid individual Cursor plans and Teams; eligible individual subscription links. Enterprise enablement is account-dependent; geographic eligibility is unverified. Grok Bot plans and billing Checked | Unknown Launched October 6 on web, iOS, and Android. Age 18 or the higher local minimum. Complete geographic eligibility is unverified. | Documented Self-hosted route needs a configured provider. Portal and Cloud need a Nous account; model and region availability depend on the route. | Unknown Manus 2.0 is announced on web, desktop, and mobile. Detailed geographic and account eligibility is unverified. Manus 2.0 and computer use Checked | Documented US and Canada, age 18+. Subscriptions require a Meta account; actual account and region offers can vary. | Unknown Provider accounts and connected services set their own access requirements. No universal regional entitlement is established. |
| Consequential unknowns | Unknown Account rollout, general BYOK, and detailed own-cloud eligibility and billing. | Unknown Model picker, BYOK, exact account deeper-work allowance, and additional-dot prices. | Unknown Exact model routing and numeric usage allowances remain unknown. Linked grants do not stack; the FAQ says the larger allowance applies. Actual account charges are untested. | Unknown Geographic eligibility, model picker/BYOK, absolute usage and reset rules, checkout currency, top-up rates, and cost per completed task. | Unknown Portal currency code, actual model/tool debits, account quotas, and the complete recurring bill. | Unknown Current models, BYOK, plan/credit selections, annual invoice totals, and Cloud Computer add-on amount. | Unknown Account-specific offers, free allowance, checkout currency, cost per completed task, model picker/BYOK, and arrival of planned Confidential VM and 1Password features. | Unknown Exact provider entitlement, complete job cost, and reliability of your chosen model/tool configuration. |
| Guide and test coverage | Not tested No dedicated setup guide. Hands-on not run. Coverage note Checked | Not tested Dots guide. Hands-on not run. Dots guide Checked | Not tested Published product overview. Hands-on not run. Product overview Checked | Not tested Hark Pro versus Muse comparison. Hands-on not run. Product overview Checked | Not tested Published product overview. Setup and billing guides planned; hands-on not run. Product overview Checked | Not tested No dedicated setup guide. Hands-on not run. Coverage note Checked | Not tested Hark Pro versus Muse comparison. Hands-on not run. Product overview Checked | Not tested Published product overview. Setup and billing guides planned; hands-on not run. Product overview Checked |
| Primary sources and read dates | Claude transition and task controls (read )Claude Cowork product (read )Claude Pro pricing (read )Claude subscription and API billing (read ) | Dots access and controls (read )OpenAI Dots launch and usage (read )ChatGPT Pro tiers (read )ChatGPT Business seats (read )OpenAI Business pricing (read ) | Grok Bot hosting and model controls (read )Grok Bot plans and billing (read )Grok Bot computer and access FAQ (read ) | Hark Pro launch (read )Using Hark (read )Hark credential security (read )Hark usage plans (read )Hark terms and permissions (read )Hark privacy policy (read ) | Hermes Agent licensing and hosting (read )Hermes Cloud billing (read )Hermes model providers and authentication (read )Nous Portal plans (read )Hermes permissions and approvals (read ) | Manus 2.0 and computer use (read )Manus credit plans (read )Manus task-credit billing (read ) | Muse launch (read )Muse security and data use (read )Muse subscription plans (read )Muse business connections and markets (read )Muse current access FAQ (read ) | OpenClaw runtime and trust boundary (read )OpenClaw provider authentication (read )OpenClaw custom and local providers (read )OpenClaw usage and costs (read )OpenClaw tool and agent permissions (read ) |
Claude is merging Cowork and chat in stages. Its model picker remains, and cloud tasks can continue while your laptop is closed. Work with local files or apps needs Claude Desktop open. General third-party model keys are not established for this personal task experience.
Claude transition and task controls; Claude Cowork product. Checked September 30, 2026.
Dots is OpenAI’s personal agent in ChatGPT, powered by GPT-6 Astra. Each dot has a cloud computer, with optional local access. The first dot comes with eligible Pro or Business Premium plans, but deeper work has an allowance. A Dots model picker and BYOK remain undocumented.
OpenAI Dots launch and usage; Dots access and controls. Checked September 30, 2026.
Grok Bot is the xAI product operated on Cursor-hosted computers. Cursor chooses the serving model; the product has no customer model picker. All Bots on one user account share files, browser sessions, and logins. Linked usage grants do not stack; the FAQ says the larger subscription allowance applies. Actual account charges have not been tested here.
Grok Bot hosting and model controls; Grok Bot computer and access FAQ; Grok Bot plans and billing. Checked September 30, 2026.
Hark Pro is a managed personal agent on web, iOS, and Android. Handoff handles cloud-browser work; Google Workspace and Microsoft 365 have direct connections. Panels are custom mini-apps, and Scheduled Tasks support recurring work. Hark chooses its serving models. Its credential vault protects logins, but that does not establish whole-product end-to-end encryption or no provider access to task data.
Hark Pro launch; Using Hark; Hark credential security; Hark privacy policy. Checked October 6, 2026.
Hermes Agent is MIT-licensed software you can run yourself. Model selection and provider credentials are configurable, including compatible local endpoints. Nous Portal credits and Hermes Cloud hosting are optional services with separate meters. Native Anthropic OAuth requires Max plus extra usage credits; it does not draw on the base Max allowance.
Hermes Agent licensing and hosting; Hermes model providers and authentication; Hermes Cloud billing. Checked September 30, 2026.
Manus 2.0 runs agent work on managed infrastructure and offers an authorized local computer-use session. Its Cloud Computer is a paid add-on for projects needing a persistent environment. Current model selection, BYOK, and that add-on’s price are unverified; older Manus 1.6 labels do not establish 2.0 restrictions.
Manus 2.0 and computer use; Manus credit plans; Manus task-credit billing. Checked September 30, 2026.
Muse is Meta’s managed personal agent with its own cloud computer, background work, and WhatsApp conversation. Meta names Muse Spark and a growing consumer and business connector catalog. Sentinel mediates permissions. The current VM allows operational provider access; Confidential VM protection is planned. Meta describes training on sanitized conversations and tool traces by default, with an opt-out.
Muse launch; Muse business connections and markets; Muse security and data use. Checked October 6, 2026.
OpenClaw’s Gateway runs on your own infrastructure. You choose provider authentication and models, including documented local-runtime options. A remote model or connected service can still receive task data. The software does not establish your operating bill: inference, optional tools, and the machine remain separate costs.
OpenClaw runtime and trust boundary; OpenClaw provider authentication; OpenClaw custom and local providers; OpenClaw usage and costs. Checked September 30, 2026.
Choose the requirement that would rule a product out before comparing plans. If you must select the model provider or connect a local endpoint, start with OpenClaw or self-hosted Hermes. Budget for the host and check the chosen authentication route; neither installation makes remote inference local. OpenClaw provider authentication; OpenClaw custom and local providers; Hermes model providers and authentication.
For a managed product, check your existing account first: can it access the runtime, connect the required sources, and permit the task? Compare Hark Pro, Muse, Dots, Grok Bot, Claude, and Manus against those requirements in the table. Reject any option whose required control remains unknown until its documentation or your account confirms it. Then try one task you can inspect before paying for more usage. This is our selection approach, not a tested performance ranking.
A subscription buys an allowance under that provider’s rules. A provider API key can add an inference bill. Search, browser, and other connected tools can add charges, and a machine running the agent adds hosting or hardware costs. Hermes Cloud explicitly bills inference and tools on top of compute and storage. Hermes Cloud billing.
Keep those amounts in separate lines of your budget. Different providers’ credits do not measure the same work, and a stopped agent can still need paid storage. An unknown charge is not zero. No complete monthly or per-task bill has been measured here.
For the model portion, use the LLM API price table and model token-cost calculator. They do not estimate the whole personal-agent bill.
First confirm the product is available to your account and region. Dots is rolling out gradually, and Claude accounts can differ during its transition. A subscription name alone does not guarantee the new task experience. Dots access and controls; Claude transition and task controls.
Then check the boundary you are granting: selected folders, connected-app permissions, a logged-in cloud browser, or local command execution. Grok Bots on one account share a computer and its logins. Creating another Bot does not isolate those credentials. Grok Bot computer and access FAQ.
Self-hosting tells you where the runtime lives. It does not tell you where every request goes: remote inference and connected services can receive data. Check the chosen model route and tools before treating a local installation as a local-only setup. OpenClaw provider authentication; Hermes model providers and authentication.
The Dots guide explains its documented plans and controls. The OpenClaw, Hermes, and Grok Bot overview discusses persistent work. Neither is a matched product test.
For model selection, the agentic model shortlist and agent benchmark results cover model and harness evidence. Those scores do not rank the personal-agent products above.
Our Dots vs Grok Bot comparison works through account eligibility, usage grants, shared computers, and action permissions.
For self-hosted setups, OpenClaw pricing separates model, tool, hosting, and operator costs and includes a blank cost worksheet. OpenClaw vs Hermes Agent compares configuration and billing responsibilities. The OpenClaw alternatives ledger covers other documented operating paths. These guides do not establish a product performance winner.
The Hark Pro vs Muse comparison examines connected apps, usage allowances, and the difference between a credential vault and protection for the whole agent. It compares current documentation, not tested task completion.
We include identifiable personal-agent products with primary documentation about how they run. Products are listed alphabetically. Each cell links to the source checked for that claim, including sources inspected for an unanswered question. “Unknown” means the checked material does not settle the field; it does not mean the feature is absent. “Not tested” describes our product-test coverage.
Current setup and billing documentation takes precedence over broad launch descriptions. Documented allowances stay separate from an untested account outcome. Prices identify the currency, unit, and billing interval; unverified amounts stay out of the table.
Read dates describe source checks, not successful onboarding or invoices. Billing, eligibility, and rollout can change after October 6, 2026. Send a source URL and the affected cell through the correction form so we can check it.
Yes. OpenClaw and Hermes document remote provider routes alongside local options. Hosting the runtime yourself does not keep remote inference or connected-app requests on your machine. Check the exact endpoint, authentication route, and tool settings before deciding what data leaves the host. OpenClaw provider authentication; Hermes model providers and authentication.
It depends on the route and allowance. Dots delegates Codex and ChatGPT Work tasks to their usual usage meters. Hermes Cloud adds hosting and storage charges separately from inference and tools. Extra usage and connected services can change the bill, so an included agent is not a measured all-in operating cost. OpenAI Dots launch and usage; Hermes Cloud billing.
This comparison has no matched product task test, so it cannot establish a performance winner. The table can help you choose a configuration by hosting, model control, billing, and permissions. Model and harness benchmarks are separate evidence; they do not measure every product’s persistent work or complete job cost.