Continue the conversation — chat opens pre-seeded with the current signal, caps, and movement.
Manus positions itself as a fully autonomous general-purpose agent ('Hands On AI') that delivers finished artifacts rather than chat answers — you set a goal and walk away. Differentiator vs. dev-specific autonomous agents (Devin, OpenHands): it is general-purpose (research, data analysis, report/website generation, computer use) rather than a codebase-aware software engineer. Built by Butterfly Effect (founded in China 2022; HQ relocated to Singapore after the April 2025 Series B). Publicly launched March 6, 2025 (invitation-only beta). Orchestrates external frontier models — primarily Anthropic Claude with fine-tuned Alibaba Qwen for sub-tasks — and trains no foundation model of its own.
The defining 2026 story is corporate, not technical: Meta announced a ~$2B acquisition in December 2025; China's NDRC opened a probe almost immediately, restricted the co-founders from leaving the country in March 2026, and on April 27, 2026 prohibited the transaction — the first time a consummated deal was ordered reversed under China's Foreign Investment Security Review (FISR) Measures. By June 2026 Meta had firewalled Manus from internal systems and an internal memo described the platform as 'sunsetting.' The standalone subscription product nonetheless remains live and processing subscriptions from Singapore.
Adoption & Proof Points
- Reported ~$125M annual revenue run rate as of December 2025 (Wikipedia, citing reporting). ~100 employees post-relocation in Singapore (about 40 of ~120 China staff relocated; most others laid off). Funding: ZhenFund seed (2023); Series A from Sequoia China/HSG and Tencent (Nov 2024); ~$75M Series B led by Benchmark at a ~$500M valuation (April 2025). The Benchmark investment drew a U.S. Treasury OISP review in May 2025 that largely faded after the Singapore relocation. No named Fortune 500 enterprise production deployments surfaced in desk research — adoption signal is consumer/prosumer subscription plus revenue run-rate, not enterprise references.
Risks & Limitations
- VIABILITY/OWNERSHIP (critical): Meta's ~$2B acquisition blocked and ordered reversed by China's NDRC (Apr 27, 2026); Meta firewalled Manus and an internal memo calls it 'sunsetting'; ownership and strategic future unresolved as of June 2026.
- Reliability: independent 2026 reviews report tasks failing mid-stream, hallucinated browser clicks, 'hallucinated progress,' branching-logic workflows derailing, and poor server reliability at peak — 'not a reliable tool for mission-critical work.'
- Cost predictability: subscription-plus-credits model where credits burn unpredictably per task; monthly and daily-refresh credits do not roll over; pricing described by reviewers as hard to forecast.
- Benchmarks: GAIA leadership claims are vendor-reported from the invitation-only era and not independently corroborated.
- Data privacy: Reddit reports of public exposure of user names/task details.
- Not dev-native: general agent without codebase indexing, native SCM/PR workflow, or IDE integration.
- Model dependency: relies on third-party Claude + Qwen; no owned foundation model.
Capabilities & Integration
Agentic depth: autonomous multi-step execution in a sandboxed environment — Chromium browser, Linux terminal (Python/bash), and a file system. Can scrape websites, download CSVs, write/run code, call REST APIs, and synthesize multi-source research into reports. Web App Builder generates full sites/apps with built-in database, Stripe, and SEO. Manus Desktop (March 2026) adds a 'My Computer' mode: reads/writes files in authorized directories, executes terminal commands behind per-command approval gates ('Allow Once' / 'Always Allow'), uses local dev environments (Python, Node.js, Swift, Xcode) and machine GPU, and can accept remote tasks to act as a persistent worker.
Model architecture: orchestrates multiple LLMs rather than a single base model — primarily Anthropic Claude, with fine-tuned Alibaba Qwen handling sub-tasks. No in-house foundation model.
Context/memory: multi-source web research and filesystem/scheduled-task persistence; no codebase indexing, semantic code search, or cross-repo awareness (not a codebase-aware coding agent).
Integration: general tool access (browser/terminal/API/filesystem) and a Web App Builder, but no documented native Git/GitHub PR workflow, no GitLab/Bitbucket, no confirmed MCP server, and no Slack/Teams/Linear/Jira agent connectors — thin on the developer-workflow integrations the Autonomous Agents cohort is measured against.