Evaluation record · manus

Manus

v1.x

Butterfly Effect / Meta (acquisition being unwound as of June 2026)

Agentautonomousgeneral-purposecloud-agentconsumer

Adequate

About This Agent

General-purpose autonomous agent that executes end-to-end tasks in a cloud VM equipped with a browser, shell, and file tools. Launched virally in March 2025 by Butterfly Effect and acquired by Meta for over $2B in late December 2025 — but Chinese regulators ordered the deal reversed in April 2026, and by June 2026 Meta had completed an operational split (halting data sharing) while it dismantles the acquisition; Manus co-founders are reportedly in talks to buy the company back.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

task completion accuracy

Assessment of end-to-end task completion based on published benchmark claims and independent user testing reports

Evidence

Wikipedia - Manus (AI agent) — Strong showing on GAIA-style autonomous task benchmarks at launch; real-world completion varies widely with task ambiguity and website complexity

mediumVerified: 2026-07-09

tool use reliability

Review of integrated tool stack reliability across browsing, shell, and file operations

Evidence

Manus Product Page — Integrated cloud VM toolset: full browser automation, shell, file system, code execution, and document/slide generation

mediumVerified: 2026-07-09

multi step planning

Evaluation of autonomous task decomposition and long-horizon asynchronous execution

Evidence

Wikipedia - Manus (AI agent) — Decomposes high-level goals into explicit task plans and executes them asynchronously, continuing after the user disconnects

mediumVerified: 2026-07-09

memory persistence

Review of session workspace persistence and cross-task preference memory

Evidence

Manus Product Page — Session workspaces persist files and context; remembers user preferences and prior instructions across tasks

mediumVerified: 2026-07-09

error recovery

Assessment of retry behavior and failure modes from independent reviews

Evidence

Wikipedia - Manus (AI agent) — Retries failed steps and adapts plans, but early reviews noted loops, stalls on complex sites, and occasional silent task failures

lowVerified: 2026-07-09

agent collaboration

Review of internal multi-agent architecture versus user-controllable collaboration

Evidence

Manus Product Page — Internally orchestrates multiple sub-agent roles (planner, executor, verifier); no user-facing multi-agent composition

lowVerified: 2026-07-09

🛡️Security

tool sandboxing

Security architecture review of cloud VM isolation model

Evidence

Manus Product Page — All execution occurs in isolated per-session cloud VMs, never on the user's device

mediumVerified: 2026-07-09

access control

Review of credential handling and account-level controls

Evidence

Manus Product Page — Consumer-oriented account model; credentials given to the agent for logged-in browsing are held in the session, with limited enterprise-grade access governance

lowVerified: 2026-07-09

prompt injection defense

Threat surface analysis of autonomous browsing with limited disclosed defenses

Evidence

Wikipedia - Manus (AI agent) — Autonomous web browsing across arbitrary sites is a large injection surface; public documentation of mitigations is minimal

lowVerified: 2026-07-09

data isolation

Data architecture review accounting for ownership and jurisdiction transition, updated for the June 2026 unwind and data-sharing cutoff

Evidence

Bloomberg - Meta severs Manus data access — Per-session VM isolation exists, but data residency and processing arrangements are in flux again: after Beijing ordered the buyout unwound, Meta completed an operational split from Manus and halted data sharing between the companies in June 2026

lowVerified: 2026-07-09

open source transparency

Source availability assessment

Evidence

Manus Product Page — Proprietary agent harness; historically orchestrated third-party models (Claude, Qwen) rather than publishing its own stack

highVerified: 2026-07-09

🔒Privacy & Compliance

data retention

Review of published retention practices during ownership transition

Evidence

Manus Privacy Policy — Session data, files, and browsing artifacts retained in Manus cloud; retention terms are consumer-grade and unsettled while the Meta acquisition is unwound

lowVerified: 2026-07-09

gdpr compliance

Compliance posture assessment during corporate separation

Evidence

CNBC - Meta begins dismantling Manus deal — Compliance ownership is unresolved: Chinese regulators ordered the $2B deal reversed (April 2026) and Meta is dismantling the acquisition, so the Meta compliance umbrella no longer reliably applies to Manus

lowVerified: 2026-07-09

third party data sharing

Data flow analysis of third-party model routing before and after acquisition

Evidence

Wikipedia - Manus (AI agent) — Historically routed tasks to third-party models (Anthropic Claude, Alibaba Qwen); under Meta, model routing and data sharing arrangements are changing

mediumVerified: 2026-07-09

local deployment option

Deployment options assessment

Evidence

Manus Product Page — Cloud-only consumer service with no self-hosted, on-premises, or offline option

highVerified: 2026-07-09

👁️Trust & Transparency

documentation quality

Documentation completeness review

Evidence

Manus Help Center — Consumer-grade guides and use-case galleries; limited technical documentation of architecture, security, or limits

mediumVerified: 2026-07-09

execution traceability

Review of live session visibility and replay features

Evidence

Manus Product Page — 'Manus's Computer' view shows the agent's browser, shell, and file actions live, and completed sessions are replayable and shareable

highVerified: 2026-07-09

decision explainability

Assessment of plan visibility and progress narration

Evidence

Manus Product Page — Displays its task plan and step-by-step progress narration, though internal model routing decisions are opaque

mediumVerified: 2026-07-09

open source code

Open source assessment

Evidence

Manus Product Page — Closed-source product; only minor components and demos have been shared publicly

highVerified: 2026-07-09

community activity

Community engagement analysis of user base and public activity

Evidence

Wikipedia - Manus (AI agent) — Viral launch 2025-03-06 with invite waitlist in the millions; open signup May 2025 built a large consumer user base and active social community

mediumVerified: 2026-07-09

⚙️Operational Excellence

ease of integration

Onboarding friction assessment for non-technical users

Evidence

Manus Product Page — Zero-setup web and mobile apps; describe a task in natural language and the cloud VM handles the rest

highVerified: 2026-07-09

scalability

Scalability assessment of concurrent task limits across tiers

Evidence

Manus Product Page — Multiple concurrent cloud tasks supported on paid tiers; throughput governed by credit balance and plan limits

mediumVerified: 2026-07-09

cost predictability

Pricing model analysis; variable credit burn per task reduces predictability

Evidence

Manus Pricing — Credit-based: Free tier with 300 daily credits; paid plans roughly $20-40/mo up to $200/mo. Credit consumption per task varies and is hard to estimate upfront

highVerified: 2026-07-09

monitoring capabilities

Monitoring features assessment

Evidence

Manus Product Page — Per-task session views and credit usage tracking; no organizational observability or audit tooling

mediumVerified: 2026-07-09

production readiness

Vendor stability and product maturity assessment; score lowered 68 to 60 to reflect the forced unwind of the Meta acquisition and unresolved future ownership

Evidence

CNBC - Meta begins dismantling Manus deal — Meta began dismantling the $2B acquisition after Beijing ordered it reversed under China's foreign investment security review; operational split and data-sharing halt completed June 2026; co-founders reportedly seeking ~$1B to buy the company back, leaving ownership, resourcing, and roadmap unresolved

mediumVerified: 2026-07-09

Strengths

+True general-purpose autonomy: cloud VM with browser, shell, and file tools executes tasks end-to-end
+Transparent 'Manus's Computer' live view and replayable sessions show exactly what the agent did
+Asynchronous execution continues in the cloud after the user disconnects
+Zero-setup web and mobile experience accessible to non-technical users
+Free tier with 300 daily credits allows meaningful evaluation before paying

Limitations

!Ownership crisis: Beijing ordered Meta's $2B acquisition reversed (April 2026); Meta completed an operational split and halted data sharing in June 2026 while co-founders seek a buyback, leaving governance, data handling, and long-term resourcing unresolved
!Minimal published security documentation; prompt injection defenses for autonomous browsing are unclear
!Credit-based pricing with variable per-task burn makes costs hard to predict
!Cloud-only with no self-hosted option; sensitive data must enter Manus's VMs
!Reliability degrades on complex or ambiguous tasks, with loops and silent failures reported
!Closed-source stack with opaque routing to underlying models (historically Claude and Qwen)

Metadata

license: Proprietary

supported models

0: Anthropic Claude (historical backend)

1: Alibaba Qwen (historical backend)

2: Model routing in flux amid the Meta acquisition unwind

programming languages

0: Natural language interface; agent writes Python, JavaScript, and shell internally

deployment type: Cloud (per-session VM with browser, shell, file tools)

tool support

0: Browser automation

1: Shell and code execution

2: File and document generation

3: Slides and website generation

4: Scheduled tasks

first release: 2025-03-06 (viral invite launch); open signup May 2025

pricing: Free (300 daily credits); paid plans ~$20-40/mo up to $200/mo (credit-based)

company milestones: Built by Butterfly Effect (Singapore HQ, Chinese origins); acquired by Meta for >$2B (closed ~2025-12-30); Chinese regulators ordered the deal unwound April 2026; Meta completed operational split and halted data sharing June 2026; co-founders reportedly raising ~$1B for a buyback

Use Case Ratings

research assistant

Flagship use case: autonomous multi-source web research, comparison shopping, screening, and report generation

data analysis

Runs code in its VM to clean data, analyze spreadsheets, and produce charts and dashboards from uploaded files

content creation

Generates documents, slide decks, and simple websites end-to-end, though output quality needs human review

code generation

Can build and deploy small apps and scripts, but lacks repository workflows and engineering rigor of dedicated coding agents

education

Builds interactive course materials and explainers; replayable sessions show how tasks are accomplished

Similar Agents

Devin

Cognition

OpenAI Codex

OpenAI

Claude Code

Anthropic

Google Jules

Google