Personal AI Agents: OpenClaw and Alternatives Explained

Personal AI agents now read your email, browse the web, run commands and act on schedules for you, yet the main products differ sharply in where they run, which models they use, what they cost and when they stop to ask permission. This page is a factual reference to four personal agents with official documentation: ChatGPT agent from OpenAI, Gemini Tasks from Google, Manus from Manus AI, and the self-hosted, open-source OpenClaw from the OpenClaw Foundation. It sets out the same fields for each: owner, deployment model, plans and price, models, platforms, capabilities, approval model and data handling. It is written for individuals choosing an agent, and for IT and security teams who need to know what staff are running. Products are listed alphabetically, and the order is never a ranking. Hardening a self-hosted agent is covered separately in the Self-Hosted Personal AI Agent Security Guide, which we do not repeat here.

Last checked: 2026-11-15. Facts come from each provider's own pages on that date. Where a provider's pages do not state a fact, the page says so.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

Two deployment models. OpenClaw runs on hardware you control and connects to the model of your choice; ChatGPT agent, Claude Cowork, Copilot Tasks, Gemini Tasks and Manus run in their provider's cloud on the provider's models.
Pricing follows the model. OpenClaw states it is free and MIT licensed, with model costs paid to whichever provider you connect; the hosted agents come with paid subscriptions (Google AI Pro or Ultra, ChatGPT paid plans) or credits (Manus).
Confirmation points are documented differently. OpenAI and Google both state their agents ask before purchases or sending messages; OpenClaw offers full-access or sandboxed operation, with exec approvals and DM pairing documented.
Chat apps are a key difference in reach. OpenClaw states it works across 29 chat channels including WhatsApp, Telegram, Slack, Signal and iMessage; the hosted agents work inside their own apps and connected services.
The agent holds your credentials. Whatever you connect (mailbox, calendar, code repositories, browser sessions) becomes reachable by the agent, so inventory and scope those connections as you would any other non-human identity.

What is a personal AI agent?

A personal AI agent is software that uses a large language model to carry out multi-step tasks for one person, such as triaging email, booking travel, monitoring prices or running scripts, often in the background or on a schedule. It differs from a chat assistant in that it acts through tools: a browser, a terminal, file access or connected apps. It differs from an enterprise agent platform in that one user, not an IT team, usually sets it up and grants its access. Agents that mainly drive a browser or desktop are covered in Computer-Use and Browser Agents; the wider category distinctions are in Agentic AI vs Generative AI.

Personal AI agents at a glance

As of 2026-11-15. Identical fields for every product; alphabetical order.

Field	ChatGPT agent	Claude Cowork	Copilot Tasks	Gemini Tasks	Manus	OpenClaw
Owner	OpenAI	Anthropic	Microsoft	Google	Manus AI	OpenClaw Foundation (US 501(c)(3) non-profit)
Launch	17 July 2025	Not stated by Anthropic on the pages checked	Not stated by Microsoft on the pages checked	Not stated by Google; Australian rollout announced 29 July 2026 under the earlier name Gemini Spark	Not stated by Manus AI on the pages checked	Not stated by the OpenClaw Foundation on the pages checked
Deployment	Hosted: runs on its own virtual computer	Hosted: runs code in an isolated environment on Anthropic's servers; reaches local files, browser or computer through Claude Desktop	Hosted: location not stated by Microsoft on the pages checked	Hosted: runs in Google's cloud, local Chrome or remote cloud browser	Hosted: web dashboard	Self-hosted on the user's machine or server
Licence	Proprietary service	Proprietary service	Proprietary service	Proprietary service	Proprietary service	MIT
Plans	Pro, Plus and Team at launch; Enterprise and Education to follow	Pro, Max, Ultra, Team and Enterprise; not on Free	Personal Microsoft accounts; migrating to Copilot Cowork	Google AI Pro or Ultra; personal accounts; age 18+	Free, Pro, Team (credits); Enterprise mentioned	Free; no subscription, hosted tier or paid edition
Models	OpenAI agentic model (not named in the launch post)	Not stated by Anthropic on the pages checked	Not stated by Microsoft on the pages checked	Gemini 3.5 (per the Australian launch post)	Not stated by Manus AI on the pages checked	User's choice: hosted, subscription-backed, gateway or local models
Platforms	ChatGPT, selected via "agent mode" in the tools dropdown	Claude Desktop (macOS, Windows, Linux beta), web, mobile, Claude in Chrome side panel	Tasks view in Copilot; platforms not stated by Microsoft on the pages checked	Gemini mobile app, Mac app, web; Chrome auto browse on desktop only	Web dashboard	macOS 15+, Windows 10/11, Linux desktop apps; Android and iOS apps; 29 chat channels
Scheduling	Recurring scheduled tasks	Scheduled tasks run in the cloud with no device online	One-off or scheduled tasks, set time or recurring	One-off or scheduled tasks; up to 15 running at once	Not stated by Manus AI on the pages checked	Automations (cron)
Regional limits	EEA and Switzerland access in progress at launch	Not stated by Anthropic on the pages checked	Not stated by Microsoft on the pages checked	Not available in the EEA, Nigeria, Switzerland or the UK	Not stated by Manus AI on the pages checked	None stated (self-hosted)

Self-hosted and hosted agents: what changes

Question	Self-hosted (OpenClaw)	Hosted (ChatGPT agent, Claude Cowork, Copilot Tasks, Gemini Tasks, Manus)
Where does state live?	On the user's machine, in local configuration and memory	In the provider's cloud, under the provider's retention settings
Who patches it?	The user, by updating releases	The provider
Who chooses the model?	The user, per provider or local runtime	The provider
Who pays for tokens?	The user, directly to the model provider	Included in the plan or credits
What sets the safety boundary?	User configuration: sandbox backend, approvals, pairing, network binding	Provider-defined confirmation points plus user settings
What does IT see?	A local process and outbound model API traffic	A SaaS service with OAuth grants to connected apps

For the wider trade-off between running models yourself and using hosted APIs, see Self-Hosted LLMs vs Hosted LLM APIs.

ChatGPT agent (OpenAI)

ChatGPT agent was introduced on 17 July 2025 and rolled out to Pro, Plus and Team users that day, with Enterprise and Education to follow. Users turn it on by choosing "agent mode" from the composer's tools dropdown.

Tools: a visual browser, a text-based browser, a terminal, direct API access and ChatGPT connectors such as Gmail and GitHub, all on a virtual computer that keeps context across tools.
Limits at launch: Pro users had 400 messages a month and other paid users 40, with more through credit-based options.
Scheduling: completed tasks can be set to recur, for example a weekly metrics report.
Approvals: OpenAI states the agent asks before consequential actions such as purchases, requires active oversight (Watch Mode) for critical tasks such as sending email, and is trained to refuse high-risk tasks such as bank transfers.
Data: data typed during takeover mode, such as passwords, is not collected or stored; one setting deletes browsing data and logs out all active website sessions. OpenAI names prompt injection as a significant risk.

Claude Cowork (Anthropic)

Anthropic's help centre describes Cowork as running code and shell commands in an isolated environment on Anthropic's servers, reaching local files, the browser or the computer through the Claude Desktop app, with scheduled tasks that run in the cloud with no device online.

Eligibility: paid plans (Pro, Max, Ultra, Team and Enterprise); Enterprise owners must enable it. Regional limits not stated on the pages checked.
Capabilities: code and shell commands in a cloud environment, files in connected folders, connectors, and browser actions through Claude in Chrome.
Concurrency: not stated by Anthropic on the pages checked.
Approvals: Manual, Auto and Skip modes; connectors set to Always allow, Needs approval or Blocked; explicit permission before permanently deleting files.
Data: deleted tasks are removed from history immediately and from backend storage within 30 days; memory can be turned off per task before the first message.

Copilot Tasks (Microsoft)

Microsoft's support article describes Copilot Tasks as Copilot browsing websites, generating or editing files and using connected services, once or on a schedule, for personal Microsoft accounts. It states that personal Copilot accounts will begin to be migrated to Copilot Cowork.

Eligibility: personal Microsoft accounts (Microsoft 365 Personal, Family, Premium and Pro). Regional limits not stated on the pages checked.
Capabilities: browser-based and non-browser tasks, file creation and editing, and connectors such as email, calendar and cloud storage.
Concurrency: not stated by Microsoft on the pages checked.
Approvals: asks for approval or hands back control for monetary transactions, submitting personal information, actions involving other people, account changes and sensitive actions; tasks do not start unless the user asks.
Data: optional cookies are encrypted and kept for up to 30 days and can be deleted; connectors and memory can be managed; Microsoft account users control training use.

Gemini Tasks (Google)

Google's help centre describes Tasks in Gemini as Gemini working on goals you describe, once or on a schedule, and states the feature was earlier tested as "Spark". Google's Australian announcement of 29 July 2026 describes Gemini Spark as a background agent on Gemini 3.5 that keeps working when the laptop is closed.

Eligibility: age 18 or over, personal Google Account, Google AI Pro or Ultra, and Keep Activity turned on. Not available in the EEA, Nigeria, Switzerland or the UK.
Capabilities: newsletter clean-up and unsubscribes, price monitoring, cited research into documents and slides, web actions in a local Chrome or remote cloud browser, and Calendar, Gmail, Docs, Sheets, Slides, Drive, Keep and Tasks management.
Concurrency: up to 15 tasks at once; schedules do not run while 15 are running.
Approvals: always asks before sending communications, changing or deleting data, purchases, submitting web forms or signing in with Google; pauses before sharing health, financial or government ID details even under "Always allow".
Data: remote browser cookies are kept for future sessions and can be deleted in Settings; turning off task mode deletes remote browser and code execution data.

Manus (Manus AI)

Manus describes itself as a full autonomous agent with a separate Chat mode for quick queries, billed on a credit system where credit use depends on task complexity.

Plans: Free (limited monthly credits), Pro (a larger monthly allocation and priority support) and Team (shared credit pool, collaboration and admin controls); the FAQ mentions Enterprise plans with shared pools.
Prices: not stated by Manus AI on the plans documentation or pricing page checked.
Tasks: multi-phase tasks with intermediate outputs; the documentation advises reviewing intermediate results on complex tasks.
Usage tracking: a web dashboard shows balance, usage history, spending insights and alerts.
Models, approvals and data handling: not stated by Manus AI on the pages checked.

OpenClaw (OpenClaw Foundation)

OpenClaw is an open-source assistant that "runs on your machine" and is used from chat apps. The OpenClaw Foundation, an independent US 501(c)(3) non-profit, employs the core team and signs releases, and keeps the product MIT licensed with "no subscription" and "no hosted tier". Desktop releases listed on 15 November 2026 were v2026.11.3 (macOS), v2026.11.1 (Windows) and v2026.11.2 (Linux).

Install: one-line installers, npm (Node.js 24.22.1+ or 26.12.1+ per the README), pnpm, bun, Docker, Nix or source; a gateway service runs the agent and a Control UI dashboard manages it.
Channels: 29 in total, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Google Chat, Microsoft Teams, Matrix and Mattermost.
Models: bring your own, from providers such as Anthropic, OpenAI and Google Gemini, or local runtimes such as Ollama.
Capabilities: persistent memory, browsing and form filling, file read and write, shell commands, community skills and plugins (ClawHub registry), cron automations and shared team gateways.
Security defaults: the gateway binds to loopback on a regular host install; most channels answer unknown DM senders with a pairing code; group access is allowlisted; tools run on the host for the main session unless sandboxing is configured (Docker, Podman, SSH, OpenShell or Crabbox backends). Exec approvals and session permission modes are documented, and openclaw security audit checks for configuration drift.
Trust model: the documentation assumes one trust boundary per gateway and states OpenClaw is not a hostile multi-tenant boundary; it advises separate gateways and credentials for mixed-trust users.
Telemetry: by default an install sends the Foundation a daily version check, which can be turned off.

How much do personal AI agents cost?

Product	Access price	Model or usage cost
ChatGPT agent	Included in ChatGPT Pro, Plus and Team plans at launch	Monthly message allowance (400 Pro, 40 other paid at launch) plus credit-based options
Claude Cowork	Included in Claude Pro, Max, Ultra, Team and Enterprise	Not stated separately by Anthropic
Copilot Tasks	Personal Microsoft accounts (Microsoft 365 Personal, Family, Premium and Pro)	Not stated separately by Microsoft
Gemini Tasks	Included in Google AI Pro or Ultra	Not stated separately by Google
Manus	Free, Pro, Team plans; prices not stated by Manus AI on the pages checked	Credits, varying with task complexity
OpenClaw	Free (MIT)	Paid to the connected model provider at its API or subscription rates; local models incur hardware cost only

A self-hosted agent that runs all day can consume many tokens. Current API rates are in the LLM Pricing Tracker; set a spend cap on every key the agent uses, as described in How to set spend limits on LLM API keys.

Approvals and permission controls

Each product's documented control, in its own terms. As of 2026-11-15.

Product	Documented confirmation points	User-side controls
ChatGPT agent	Purchases and other consequential actions; Watch Mode for critical tasks such as email; refuses high-risk tasks such as bank transfers	Takeover mode for logins; delete browsing data and log out of sessions in settings
Claude Cowork	Manual mode asks for approval; permission needed before permanently deleting files	Manual, Auto or Skip modes; per-connector Always allow, Needs approval or Blocked; admin control of Auto mode
Copilot Tasks	Monetary transactions, submitting personal information, actions involving other people, account changes, sensitive actions	Interrupt a running task; take control of the browser; unlink connectors; manage memory
Gemini Tasks	Sending communications, changing or deleting data, purchases, form submission, Sign in with Google; sensitive data sharing	Choose connected apps; delete remote browser data in Settings; turn off task mode
Manus	Not stated by Manus AI on the pages checked	Review intermediate results; usage alerts on the dashboard
OpenClaw	Exec approvals and session permission modes as configured; DM pairing for unknown senders	Sandbox backend choice, group allowlists, loopback binding, security audit command

Approval design for agents in general is covered in AI Agent Action Approvals and Spending Limits Guide.

Data handling and security

Connected accounts are delegated access: an agent with Gmail, Drive or GitHub access acts with the user's permissions. Treat each grant as a non-human identity with an owner and a review date. See AI Agent Email Access Security.
Prompt injection travels through content: OpenAI names it a significant risk and OpenClaw's README says to "treat inbound messages as untrusted input". See OWASP LLM Top 10 LLM01 Prompt Injection Guide.
Skills and plugins are code: a skill runs with the agent's access, so install only from publishers you have checked. See Agent Skills Security Guide.
Self-hosted means self-secured: network exposure, sandboxing and updates are the operator's job. See the Self-Hosted Personal AI Agent Security Guide.
Work use needs policy: Gemini Tasks currently supports personal accounts only, and staff may connect work data to personal agents. See Blocking Personal ChatGPT, Claude and Gemini Accounts at Work.

How to choose a personal AI agent: factual criteria

Where data may live: if state must stay on your own hardware, only a self-hosted agent meets that; otherwise check each provider's retention settings.
Region and account type: Gemini Tasks is unavailable in the EEA, Nigeria, Switzerland and the UK and needs a personal account; check the others for your region.
Model choice: OpenClaw lets you pick or change the model; the hosted agents use their provider's models.
Where you want to talk to it: chat channels (OpenClaw) or the provider's own app (ChatGPT, Claude, Copilot, Gemini, Manus).
Who maintains it: self-hosting means you apply updates and configure sandboxing; hosted means the provider does.
Cost shape: fixed subscription, credits, or pay-per-token to your model provider.

Practitioner checklist

List every personal agent in use and the accounts, mailboxes and repositories each can reach.
Connect only the apps a task needs, and remove grants when the task ends.
For self-hosted agents, keep the gateway on loopback or behind authenticated remote access, and run the built-in security audit after each change.
Turn on sandboxing for tool execution, and require approval for shell commands and outbound messages.
Set spend caps and alerts on every model API key an agent uses.
Install skills and plugins only from checked publishers, and review them before granting access.
Delete stored browser sessions and cookies after sensitive tasks, using each product's documented setting.
Re-check plans, regions and approval behaviour against provider pages monthly; these products change quickly.

Personal AI Agents: OpenClaw, ChatGPT Agent, Gemini Tasks and Manus Explained