Which agent fits your needs better?
Put up to four candidates side by side. Compare differences in setup, costs, models or memory, then open a profile for conditions and sources.
Memory and conversation continuity: fundamentals →
You are comparing different types of tools and access methods. General-purpose agents, coding agents, and platforms each have their own tasks and requirements.
What matters for your task
Manus
A good fit if you want research and presentations produced in a cloud workspace
Claude (including Cowork)
A good fit if you want to hand over file work and recurring reports on a paid Claude plan
Perplexity Computer
A good fit if you already use a Perplexity Pro, Max or Enterprise plan and want to use Computer within it
What to compare first
| Property | Manus | Claude (including Cowork) | Perplexity Computer |
|---|---|---|---|
| Product family | Manus | Claude | Perplexity Computer |
| Profile type | General-purpose agent and workspace | General-purpose agent in Claude | Orchestrating digital work agent |
| Tool category | General-purpose agents | General-purpose agents | General-purpose agents |
| Purpose | For research, presentations, and digital work products in a cloud workspace. | General agent work in Claude with files, connectors, skills and background tasks. Docs, Slides and Design are generally available; Dashboards and Motion extend the offering in beta. | Digital worker for tasks across real user interfaces and connected services. |
| Features | Multi-step tasks Automations Persistent cloud computer Documents and spreadsheets Presentations Websites and code Video editing Computer control Develop games Remote control from a phone | Multi-step work in the background Documents and presentations Design in the conversation Skills for recurring tasks Connectors for work context Recurring tasks Questions and approvals Live dashboards from connected data Animated explainers with Motion Edit Google files | Teammates for ongoing work Fully local Mac mode Tasks with multiple outputs Split tasks Parallel research Use connected context Customize skills Create reports and presentations Ongoing tasks and monitoring App and web access Brain Research Preview |
| Usable AI models | Models not stated: In the standard modes Manus chooses the model itself; the checked pages do not name the language models used there. Only the beta Manus Flex lets you choose provider and model, with your own key at OpenAI, Anthropic, Google AI Studio, xAI, Meta, OpenRouter, Fireworks or Modal; custom endpoints and self-hosted models are not possible. Image, video, audio and 3D still use Manus models, and which plans include Flex is not confirmed. | Own models only: The checked pages name only Claude models for Cowork, without expressly ruling out other providers. You use them through a Claude account or your own account with Bedrock, Google Cloud, Microsoft Foundry, the Claude API or a self-run gateway. In that setup, the configuration decides what the model picker offers. | Fixed choice across providers: With the effort levels, Computer picks the model itself; under "Custom" you choose one from the list. Custom uses credits, and company admins can disable models. The release notes name GPT, Claude, Grok and DeepSeek models among others, but the checked pages contain no complete list. Portable Computer also offers local models on suitable Windows PCs, and the pages do not mention own API keys. On supported Macs, fully local mode offers Qwen 3.6 and PPLX Pro at launch. |
| Environment | Web, Manus Studio, mobile app and optional Cloud Computer | Claude on web, desktop, and mobile; gradual rollout | Web, mobile and desktop. Fully local on supported Apple silicon Macs. |
| Setup | Account and task description | Account, selected files, and connections | Perplexity account and eligible plan. Set up a model in the app for local Mac mode. |
| Connections | Browser, email, and Slack | Files, browser, connectors and Google Workspace; data platforms for dashboards | Slack and Microsoft 365 |
| Account configuration / review date | Manus 2.0 on the web with a personal account or Team plan and connected connectors (2026-11-05) | Claude Cowork on a paid plan (Pro, Max, Ultra, Team, Enterprise), desktop app, web and mobile (2026-11-05) | Perplexity Computer with a Pro or Max account or an Enterprise organisation (Enterprise Pro/Max) (2026-11-05) |
| Account used | Documented: Integrations use OAuth 2.0 or API keys, and MCP connectors use OAuth. The connections are personal to your account. | Documented: Sessions and files live with your Claude account and follow you across desktop, web and mobile. Cowork reaches local files, the browser and computer control through the desktop app; only the Chrome side panel can start a session without it. | Documented with conditions: Computer is available on Pro, Max, Enterprise Pro and Enterprise Max; in Enterprise, access depends on seat, plan type and the Computer permission. The reviewed pages do not describe the identity Computer uses in connected services. Custom API credentials have their own help article, which was not opened. |
| Admin approval | Documented with conditions: The provider's pricing page lists internal access control, SSO and usage analytics for the Team plan. The reviewed pages do not describe how admins approve or block connectors. | Documented with conditions: In Team and Enterprise, Cowork is available where Chat and Cowork unified is turned on. Owners can turn Cowork off for the organisation, turn off web search and manage Claude in Chrome; admins control whether the Automatically approve mode can be used. Groups and custom roles can enable Cowork for specific teams. The product page says admins manage access by team and spend limits. | Documented with conditions: On Enterprise, admins can turn Computer off for the whole organisation, disable connectors, and limit an enabled connector to specific roles. The help page recommends Disabled as the default for new connectors. Role-based connector control is available on sales-assisted Enterprise plans according to the help page. |
| Data access & team sharing | Documented: Connections are not shared with other users unless you explicitly collaborate on a task. A Cloud Computer is isolated from your own machine, and permissions depend on who sent the last message in a session. | Documented: Claude can only read and write files in folders you have connected. Through connected apps it can also read email and documents. Sessions cannot be shared, but individual artifacts can. | Documented: According to the product page, Computer connects to Gmail, Slack, Notion, Calendar and other tools. For Microsoft 365 connectors, Perplexity says existing enterprise policies for access, retention and audit are inherited. The role check runs server-side each time a connector is used, including automations. Permissions can additionally be set per tool. |
| Deployment covered | Manus 2.0 with scheduled tasks, automations and an optional Cloud Computer | Cowork cloud sessions and local tasks through the desktop app | Hosted Computer sessions in isolated containers with automations, on web, desktop and mobile |
| Work state retained | On a Cloud Computer, created files and tools installed by Manus remain, and every task shares the same hard drive. If the subscription lapses, the virtual machine is shut down and working files, tools and databases are deleted; delivered files stay in the chat history. | Claude's work runs in an isolated environment on Anthropic's servers. Sessions and files stay with the account; deleted tasks leave the history immediately and are deleted from backend storage systems within 30 days. | Each task session runs in its own isolated compute container with a dedicated filesystem and browser. Computer has persistent memory across sessions and platforms. The page does not say how long files persist; Computer follows the organisation's data retention settings. |
| Starting further runs | Scheduled tasks start work at set times. Automations can also start from events in connected services, such as a new email, a calendar event, a Slack message or a Notion update. | You create recurring tasks with /schedule or the Scheduled sidebar; they do not need your computer to be awake. Tasks that use local files, the browser or the computer require the desktop app to stay open. | Computer runs tasks in the background. It can monitor email, calendars, flight status and files with condition-based triggers and run scheduled jobs such as morning briefings. Automations carry their own model selection; if that model is unavailable, the default model is used. |
| Oversight (assessment) | Editorial inference: Scheduled and event-driven runs are possible, while failure display and approvals remain undocumented. With a Cloud Computer, plan your own backups, since files are lost on cancellation. | Editorial inference: The difference between cloud and local runs decides whether tasks can run unattended. For scheduled tasks without a documented failure path, it is worth checking results yourself regularly. | Editorial inference: Permissions, audit log and limits are documented, whereas an approval flow and failure notifications are not. Oversight mainly means setting connectors and credit limits deliberately and checking results yourself. |
| Memory: scope / sources accessed | Manus Projects in Manus 2.0 (2026-11-04) | Claude Chat and cloud Cowork; excluding local Cowork sessions (2026-11-04) | Computer with Brain; Max/Enterprise Max research preview (2026-11-04) |
| Storage and scope | Partly documented: Projects retain instructions, files and sources in project context; physical storage location and backend are not named on the product page. | Partly documented: Claude saves individual memory topics during conversations in account or project memory. Physical storage location and backend are not published. | Documented: Brain maintains a context graph as an LLM wiki loaded into the agent sandbox. |
| Selection and timing | Partly documented: New tasks use project configuration; automatic distillation of all prior conversations is not established. | Documented: Claude saves automatically or following an explicit remember request. | Documented: Brain updates the wiki overnight from sessions, connector results, document changes and corrections. |
| Inspect, edit and delete | Partly documented: Project members with edit access can maintain shared instructions and files. This page does not describe individual learned memories. | Documented: Memories can be reviewed and edited; pause preserves them, reset deletes them including project memories. | Partly documented: Entries link to their originating session, file or source. Individual editing and deletion are not explained in the announcement. |
| Agent and channel boundaries | Documented: Context belongs to the project; individual tasks remain private until separately shared. | Documented: Cloud Cowork and Chat share memory. Projects can use account memory or be isolated through Use account memory. | Partly documented: Brain is account-private by default; projects can use their own context or include personal sessions. |
| Retrieval | Documented: New project tasks use saved instructions and sources. Full availability of previous conversations is explicitly not promised. | Documented: Memory complements separate RAG search of earlier chats; project searches stay within the respective project. | Documented: Computer traverses linked wiki pages and uses previous successful approaches and corrections. |
| Session reset and compaction | Partly documented: New tasks can reuse project context. Session compaction and deletion of learned memories are not described here. | Partly documented: New regular chats can use account and project memory; memory reset deletes saved memories including project memory. This memory guide does not describe a procedure for compacting long session contexts. | Partly documented: The wiki supplies prior work context to new tasks; a session-reset or compaction contract is not published. |
| Pricing model | Free $0/month; Pro from $20/month; Team from $20/seat/month. Cloud Computer costs extra. | Free: USD 0; Pro: USD 20/month or USD 200/year; Max: USD 100 or 200/month | Computer uses plan-specific included usage and credits; 100 credits currently equal $1. Max costs $200/month. |
| Limits & control | Persistent servers require a Cloud Computer subscription. Back up working files before it ends; authorize local computer control explicitly. | Dashboards is in beta on paid plans; Motion on Team and Enterprise. Team accounts require administrators to enable Google Workspace connectors. | Plan credits and administrator-approved connectors determine available work. Configure Microsoft 365 plugins and permissions for the organization. |
| Details & sources | View profile → Memory: scope / sources accessed → Sources in profile → Provider source ↗ | View profile → Memory: scope / sources accessed → Sources in profile → Provider source ↗ | View profile → Memory: scope / sources accessed → Sources in profile → Provider source ↗ |
The overview summarizes provider documentation. Cost models are categories, not current price offers. Sources and further details are available in each profile.
What matters when choosing an agent?