Manus

Short answer. Manus is a general-purpose autonomous AI agent by Butterfly Effect (the company behind Monica.im) — an “action engine” that executes whole tasks end-to-end, automates workflows and operates a browser, sold directly to consumers and teams as a cloud subscription.

Key facts.

• Autonomous multi-step task execution with a browser operator and tool use — the flagship tasks are creating slides, building websites and simple games, video and design.
• Manus 2.0 (2026) reframed the product around “less structure, more intelligence,” adding Wide Research, Mail Manus and Slack integration.
• Distributed via web, desktop and mobile clients, with a Team plan (SSO), a Startup program and a developer API at open.manus.ai.
• Cloud-only, proprietary, subscription-billed.
• Builds no foundation models of its own — it orchestrates third-party models; the official site does not name them (third-party sources describe Claude and Qwen routing).
• Butterfly Effect was founded in China and is based in Singapore, also known for the Monica assistant.

What this means. Manus is the clearest test of whether a general autonomous agent can be a standalone product — detached from any single model vendor, competing on doing whole tasks rather than on a model, benchmark or framework.

What is uncertain. The underlying model orchestration (which third-party models it routes to), exact subscription tiers (client-side rendered), the company’s funding, and independent evaluation of its task-completion quality are not documented.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-manus-1	Manus official site	https://manus.im/	Official	—	2026-09-30	high	—
src-agents-manus-2	Manus pricing	https://manus.im/pricing	Official	—	2026-09-30	high	—
src-agents-manus-3	Manus API documentation	https://open.manus.ai/docs	Official documentation	—	2026-09-30	high	—
src-agents-manus-4	Manus (AI agent) — Wikipedia	https://en.wikipedia.org/wiki/Manus_(AI_agent)	Literature	—	2026-09-30	medium	—

Why it matters

Manus matters because it is the agent itself, sold directly to end users. Every other agent in this database is either bound to a lab’s models (Kimi Code, GLM Coding Plan) or is a platform/framework for building agents (Dify, MetaGPT); Manus is the product.

China AI Hub analysis indicates Manus’s structural role is demand creation: it demonstrated, at viral scale, that consumers and knowledge workers will pay for a general “do it for me” agent, which raised the bar for every other agent product to ship autonomous, end-to-end task completion rather than chat-plus-tools. That is a product-velocity bet, not a model bet.

China AI Hub analysis: Manus is the clearest case of the agent-decoupled-from-model bet — it rises or falls on execution quality and product design, not on a model it owns. The trade-off is that its capabilities are bounded by whatever third-party models it can route to, a supply-chain dependency the official site leaves unstated. Its dependence on undisclosed models is simultaneously its flexibility and its risk.

How it differs from the vertical coding agents

Manus is a general agent, not a coding agent, and the contrast sharpens its position.

• Kimi Code and Qwen Code are coding agents bound (or, for Qwen Code, closely aligned) to their vendors’ models, optimized for whole-repo software work.
• Trae is ByteDance’s AI-native IDE, splitting a coding agent (TraeCode) from a work assistant (TraeWork).
• Manus is general: it builds slides, websites, games and video, runs research and automates email/Slack — not a code editor, but an “action engine” that happens to code when a task needs it.

China AI Hub analysis: the difference is scope-versus-depth. The vertical coding agents trade breadth for depth in software work and inherit a vendor’s model ceiling; Manus trades depth-in-coding for breadth-across-tasks and inherits a portfolio of undisclosed third-party models instead of one vendor’s. A buyer choosing Manus is choosing a general executor; a buyer choosing a coding agent is choosing a specialist bound to a known model.

Practical implications

For consumers and knowledge workers. Manus is the “do it for me” surface — slides, websites, research, email/Slack automation — sold as a subscription. The cost is cloud-only lock-in and an undisclosed model supply chain.

For teams. The Team plan (SSO) and Startup program are the managed adoption path; the developer API at open.manus.ai extends the agent to programmatic use, though its boundary with the consumer product is not fully specified publicly.

For evaluators. Manus cannot be assessed against a model benchmark because it builds no model — it must be evaluated on end-to-end task completion, and no independent evaluation of that is recorded in this database.

What the evidence shows

The evidence is strong on Manus’s product surface and weak on its internals. The official site is explicit about the flagship tasks (slides, websites, games, video, design, browser operator, Wide Research, Mail Manus, Slack) and the Manus 2.0 “less structure, more intelligence” reframing, which grounds the “general executor” reading in primary material. The company’s China-founded, Singapore-based identity and its Monica.im connection are documented in the Wikipedia entry.

The internals are the gap. The underlying model orchestration is not named on the official site — third-party sources describe Claude and Qwen routing, which the database records at medium confidence and does not treat as official fact. Subscription pricing is client-side rendered and was not extracted. China AI Hub analysis indicates this opacity is structural to the agent-decoupled-from-model bet: a product whose value is execution, not a model, has every incentive to document its tasks and little incentive to document its model supply chain, which it may not fully control.

Where this fits

Workload	Relevance
Autonomous end-to-end task execution	High
Slides, website, game and video creation	High
Browser automation via Manus browser operator	High
Deep research (Wide Research)	High
Email / Slack workflow automation	High
Whole-repo software development	Moderate (codes when a task needs it, not a coding IDE)
Self-hosted / on-premises deployment	None (cloud-only, proprietary)

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

Field reference

Field	Value	Evidence type
Underlying model	Not publicly documented (orchestrates third-party models)	Not publicly documented
Target users	Consumers, knowledge workers, teams	Official
Platform	Web app, desktop app, mobile app	Official
OS	Cross-platform (web + desktop + mobile)	Official
Browser / computer use	Yes (Manus browser operator)	Vendor-reported
Coding	Yes (builds websites and games)	Vendor-reported
Autonomous task execution	Yes (core capability)	Vendor-reported
MCP	Not publicly documented	Not publicly documented
Tool calling	Yes (browser operator, tools)	Official
Memory	Yes (task context and workspace)	Vendor-reported
Workflow	Autonomous multi-step task execution	Vendor-reported
API	Yes (open.manus.ai)	Official
Pricing	Subscription-based tiers (manus.im/pricing)	Vendor-reported
Region	Global	Official
Open-source	No — proprietary	Official
Deployment	Cloud	Official
Limitations	Underlying models undisclosed; cloud-only; builds no own models	Official
Source	manus.im · pricing · API	Official
Last verified	2026-09-30	Official

See the Butterfly Effect company profile, the choosing-an-agent guide, and the site’s AI agents and computer use technology pages. For the structural reading of Manus as the general agent against the vertical coding agents, see the research on the agent ecosystem structure and the rise of Chinese AI agents.

Labels used above: Official fact (from the Manus site, pricing and API docs), Vendor-reported claim (capability statements by Butterfly Effect), Third-party (the Wikipedia entry on the company’s provenance and third-party model-routing descriptions), and China AI Hub analysis (our synthesis, always introduced as such). No independent evaluation of Manus’s task-completion quality is currently recorded.