What Manus is
Manus is an AI agent platform for research, file creation, application development and work automation. It combines models, tools and execution environments to progress through multi-step tasks and produce reviewable results.
Specify a verifiable outcome: the file required, sources, audience and constraints. A conversation may end with an answer; agent work can also leave documents, code and actions in connected services. Assess the whole process, including intermediate decisions and work that could not be completed.
Reference documentation: Manus: catálogo de funciones.
Butterfly Effect, Monica and the early milestones
Official documentation identifies Manus and Monica as Butterfly Effect products. Monica is an earlier product within the same company: Paddle’s customer story, featuring its team, dates its launch to March 2023 and describes a platform combining multiple models for assistance tasks.
In a retrospective signed on 26 June 2025, the Manus team placed the agent’s launch three months earlier. This distinguishes the Monica product of 2023 from Manus’s emergence in March 2025.
The December 2025 corporate announcement confirmed that the company operated from Singapore and would continue to do so. This documents its operational base at that time; it does not by itself establish the exact relocation date.
Reference documentation: Monica: FAQ oficial sobre la relación con Manus y Butterfly Effect · Paddle: historia de cliente con el equipo de Monica AI · Manus Team: retrospectiva de los tres primeros meses, 26 de junio de 2025 · Manus: incorporación a Meta, diciembre de 2025.
Origins and company history
Manus emerged in 2025 as a general-purpose agent. Its March 2026 annual letter describes a progression from cloud tasks to complete applications. Xiao Hong, also known as Red Xiao, is identified as CEO in the December 2025 announcement; the founding team includes Peak Ji and Tao Zhang.
Chetan Puttagunta confirmed Benchmark invested before launch and that he served on the board. Distinguish investment, subsequent corporate transactions and the customer-facing product. A financial valuation or demonstration does not establish reliability on a real task.
Reference documentation: Manus: primera carta anual · Chetan Puttagunta: inversión de Benchmark y equipo de Manus · Manus: incorporación a Meta, diciembre de 2025.
Meta, regulatory review and renewed independence
Manus announced it was joining Meta on 29 December 2025. On 27 April 2026 China’s NDRC prohibited the foreign investment in the acquisition and required its reversal. On 1 September 2026 Manus announced formally resumed independent operations under its founding team.
Some accounts required backup and restoration and experienced temporary interruption during separation. The company attributed this to regulatory requirements, not a security breach. This is a bounded historical episode, not an assertion that all current accounts are affected.
Organisations should retain important outputs outside a single platform and test recovery. Downloaded files, shared tasks and executable projects require different checks.
Reference documentation: Manus: incorporación a Meta, diciembre de 2025 · NDRC: decisión sobre la adquisición de Manus, abril de 2026 · Manus: regreso a operaciones independientes, septiembre de 2026 · Manus: efectos de la transición sobre datos de algunas cuentas.
Manus 2.0, Cascade and Cue
Version 2.0, announced on 28 September 2026, introduced Cascade, Manus Studio and event-triggered automations. It also presented Cue as a separate personal-agent application. Do not conflate the products’ features or access conditions.
Record the environment and configuration tested when comparing versions. Successful demonstrations do not mean every capability is included in every account. Vendor improvement figures depend on the experiment and do not guarantee organisational savings.
Reference documentation: Manus: lanzamiento de la versión 2.0.
Planning and reviewing the brief
Plan Mode provides a reviewable document of goals, steps and constraints before execution. It is explicitly activated and can be introduced during a task to replan the next phase.
For OSH, separate retrieval, interpretation and production. Identify accepted sources, decisions retained by the practitioner and excluded data. Update the plan when the purpose changes so stale instructions do not guide subsequent decisions.
Approval of an outline should not imply general permission to send, publish or change records. Explicit boundaries make the resulting work easier to review.
Reference documentation: Manus: Plan Mode.
Wide Research and source comparison
Wide Research distributes comparable items among parallel agents and combines their findings. Each subtask can investigate with separate context; synthesis requires consistent criteria.
Compare OSH guidance using publisher, date, jurisdiction, audience, scope and link. Mark missing dates and inaccessible sources without inferring content from titles.
Review a sample before scaling. Check duplicates and translations of the same document: counting them as independent evidence would exaggerate coverage and could distort the final selection.
Reference documentation: Manus: Wide Research.
Browser Operator and authorised sessions
Browser Operator connects Manus to the local browser through an extension and authorised existing sessions. It differs from cloud browsing, and access can be revoked.
An open session may allow both reading and modification. Specify read-only work, drafting or saving. Content on a visited page is not a user-authorised instruction.
Start an OSH pilot with public documents. For internal systems, restrict the account and inspect actions. A description of an attempted change does not prove it succeeded or affected the intended record.
Reference documentation: Manus: Browser Operator.
Local computer and Cloud Computer
My Computer, introduced for macOS and Windows in March 2026, works with authorised folders and command approvals. Cloud Computer provides a persistent environment distinct from a task’s temporary workspace.
A one-off report need not require a permanent server. Continuous processes need backups, ownership and recovery plans. Technical persistence is not guaranteed records retention.
Cloud Computer documentation says stopping payment closes the environment and deletes working files. Separate delivered outputs from databases, programs and dependencies required to reproduce them. Test exit arrangements before depending on one installation.
Reference documentation: Manus: My Computer · Manus: Cloud Computer.
Projects and knowledge continuity
Projects groups instructions, files and tools. Sharing context does not automatically share each individual task. Importing from Drive does not create continuous synchronisation.
In OSH, identify reference versions and validity. When a procedure changes, check which tasks use the new source. A response consistent with an old file may still be wrong for current work.
Assign responsibility for maintaining instructions and retiring obsolete documents. Project memory should support continuity without turning conversational assumptions into permanent, unreviewed rules.
Reference documentation: Manus: Projects.
Skills, connectors and automations
Skills bundle reusable instructions and may include scripts. Connectors and MCP servers provide external service access. A workflow recipe does not itself grant access to required data.
An OSH automation could draft periodic summaries of approved sources. Define the period, exclusions, failure handling and reviewer. When nothing changes, report no verified updates rather than inventing material to fill a template.
Test missing sources, revoked permissions and duplicate results. Review external Skills before execution, especially code, downloads or write actions.
Reference documentation: Manus: Agent Skills · Manus: conectores y MCP.
Documents and editable presentations
PowerPoint mode, introduced in July 2026, generates native PPTX with editable charts and tables. Its announcement described a beta; verify availability for the account and mode used.
For OSH indicators, inspect embedded data as well as visible figures. An aggregated slide may retain a more detailed underlying table. Check denominators, units, period, notes and sources.
Open the final file in the recipient’s application. Verify editability, clipping and consistency with the reviewed version. A correct preview does not establish full compatibility.
Reference documentation: Manus: presentaciones PowerPoint nativas.
Video, applications and games
Video Editor in Manus Studio for macOS and Windows retains editable timeline layers. Game Dev extends game creation; its October 2026 announcement acknowledges aspects of 2.0 still needing refinement.
In OSH training, an interactive prototype can rehearse decisions. Define assessment, feedback and excluded situations. Completing a game does not establish competence for real work.
Review equipment, protective measures and movements in every video scene. Convincing imagery can depict incorrect practice. Retain scripts, sources and originals so corrections preserve content traceability.
Reference documentation: Manus: Video Editor · Manus: Game Dev.
API, integration and Manus Flex
API v2 manages tasks, files, projects and notifications of completion or needed input. Documentation marks v1 deprecated. Flex connects inference providers using your own key: providers bill inference while other services still consume Manus credits.
Define task identifiers, results and failure handling. Retrying without checking state can duplicate records or sends. Separate read and write operations for different review requirements.
Use fictional OSH data for pilots and test permissions, failures and recovery. Minimise technical logs and protect keys outside shared documents.
Reference documentation: Manus: documentación API v2 · Manus: Flex y proveedores de inferencia.
Plans, credits and usage cost
Manus distinguishes Free, Pro and Team. Consumption depends on models, environments and services; monthly and additional credits have different renewal and validity rules.
Compare cost per accepted task, including retries, corrections and review. A bounded assignment helps estimate usage before scaling to many documents.
Account for persistent infrastructure, connectors and external providers. A credit balance does not explain the full cost of operating and maintaining a solution.
Reference documentation: Manus: consumo y vigencia de créditos.
Privacy, testing and output quality
Before entering workplace information, review account terms, connected services, recipients and retention. Permissions and written instructions serve different purposes. A confidentiality instruction does not amend a provider’s contract.
GAIA is a research benchmark for general assistants combining reasoning and tools. Performance on such an evaluation does not establish competence for every OSH process.
Test representative cases with verifiable evidence. Measure omissions, incorrect citations, unintended operations and recovery. OSH quality also requires preserving context, jurisdiction and professional responsibility.
Reference documentation: GAIA: investigación original sobre evaluación de asistentes.
Applications in occupational safety and health
It can help with preparation tasks that gather information and turn it into useful prevention material. An assignment should specify permitted sources, scope and approval points, particularly when it involves browsing, connected tools or potential actions beyond preparing a draft. This makes the task boundaries clear before the agent begins working.
In OSH, it can prepare an inventory of public institutional guides, compare training materials or gather documented characteristics of communication resources. Another application is turning checked findings into a presentation draft. Sabentis proposes these scenarios for preparatory work: final selection depends on the intended population, workplace context and reliability of the sources.
Worked example and checks
For example, a team requests a comparison of public training resources about occupational road safety. It defines review criteria, asks for source links and manually checks the currency and intended audience of each resource before choosing its final selection.
- Define a bounded question: public resources for a work-related road safety session, with desired languages and audience.
- Create comparable fields: organization, title, URL, visible date, format, stated duration and conditions of use.
- Require undocumented information to be marked, without estimating durations or filling gaps by analogy.
- Review a small initial group to check that every field has the same meaning.
- Expand the search and verify links, duplicates, versions and a sample of each resource type.
The result would be an evidence-backed shortlist, rather than an exhaustive list of everything available. A page without a date may be retained with that limitation noted. A resource about private driving should not automatically be classified as specifically work-related material. Before recommending an item, open it and check that its content matches the description, not just the title. Record which entries were checked directly and which still require review.
Limitations and conditions of use
An agent may complete steps without every intermediate decision being obvious. Review the process and sources as well as the final deliverable. A comparison table does not prove exhaustive coverage or suitability for every organization.
Parallel execution does not correct ambiguous criteria: it can multiply the same confusion. Websites change, require access or prevent tasks from completing. Browser Operator acknowledges limitations with complex interactions. Where authenticated systems are involved, define reading, preparation and modification clearly; publishing, sending or changing records requires specific authorization within the workflow.
Choosing and complementing the tool
Choose Manus when work can be divided into comparable tasks with a verifiable deliverable. Excel helps review the matrix and filter candidates; Perplexity can complement targeted source-backed questions. For one simple question, coordinating multiple subtasks may add unnecessary complexity.
The initial trial should measure field quality, traceability and handling of missing information. Counting generated rows does not establish quality: a long table may contain repeated versions or inconsistent interpretations. Retaining the criteria makes it possible to repeat the review when resources change. Keep the original links available even after turning the comparison into slides or a summary.
Frequently asked questions
Does Wide Research guarantee complete research?
No. It expands parallel work, but coverage depends on sources, criteria, access and verification.
Does Browser Operator always use the cloud browser?
No. Documentation distinguishes the cloud browser from the extension operating in an authorized local browser.
Can an agent act in all my accounts?
It should work only within authorized access and scope. An open session does not itself define which actions have been approved.
