Manus

Manus is an autonomous AI agent built by Singapore-based Butterfly Effect, the company behind Monica.im, and publicly launched in March 2025 after a viral invite-only debut. In December 2025, Meta acquired it for over $2 billion, making Manus one of the most expensive autonomous-agent acquisitions in AI history. The product accepts a plain-English task, then executes it end-to-end inside a cloud-hosted virtual computer: opening browsers, clicking through web pages, downloading files, writing code, and exporting results, without requiring the user to write a single line of code or set up any API. As of April 2026, it operates as a standalone subscription product under Meta's ownership, with public signup open to anyone.

Manus handles browser automation, parallel task processing, file upload and download, real-time screen streaming so you can watch it work, natural-language command execution, and data export to CSV or Google Sheets. Under the hood, it runs on a multi-agent orchestration layer that calls Anthropic's Claude (Claude Sonnet 3.7 at launch) and fine-tuned versions of Alibaba's Qwen for specific subtasks, coordinating roughly 29 specialized tools through a browser-use framework. Users assign a task and can walk away; Manus runs asynchronously in the cloud and notifies when done. The platform claimed an 86.5% score on the GAIA Level 1 benchmark in March 2025, ahead of OpenAI's 74.3% at the time.

What Manus actually does in April 2026

Manus is a general-purpose autonomous agent, not a chatbot. You give it a goal: "Pull pricing from these 40 competitor sites and organize into a spreadsheet, flagging the lowest prices." Manus opens a virtual browser, navigates each site, extracts data, formats it, and returns a complete file. You do not supervise each step. The agent supports up to 20 concurrent tasks on paid plans, runs scheduled tasks on a timer, and maintains context across long multi-step workflows until either the task completes or the session times out.

The platform supports three autonomy modes: fully autonomous (set it and walk away), checkpoint-based (it pauses at sensitive steps for your approval), and supervised (you confirm each action). This makes it more configurable than competitors that interrupt constantly. Manus can generate slide decks, build simple web apps through iteration, conduct wide research across dozens of sources, and handle file transfers between services. The Customizable ($40/month) plan adds Wide Research and website generation features not available on the base Standard tier.

By December 2025, Manus had processed over 147 trillion tokens and created more than 80 million virtual computers, reaching $100 million in ARR approximately eight months after launching paid plans. That velocity attracted the Meta acquisition. Post-acquisition, the subscription service continues operating from Singapore with no announced product changes beyond eventual integration into Meta's platforms.

"Manus has been the best agent for turning text into work -- useable slides, code, research. Nothing else comes close for end-to-end delegation." -- HN user, Hacker News, December 2025

Where Manus sits versus Devin and OpenAI Operator

Devin (Cognition AI) is purpose-built for software engineering. It provisions a complete development environment: IDE, browser, terminal, and direct filesystem access. Devin writes code, runs tests, debugs, and deploys via CLI commands rather than clicking through browser UIs. Its mechanical advantage over Manus is direct code execution without the browser automation bottleneck, making it significantly faster and more reliable for pure software tasks. The tradeoff: Devin cannot handle the 90% of knowledge-work tasks that do not involve writing code. Its official SWE-bench success rate is 13.86%. Manus can do market research, format spreadsheets, generate presentations, and manage files across services. Devin cannot.

OpenAI Operator uses the same browser-automation model as Manus (controlling a browser via CDP/Selenium, clicking DOM elements rather than calling APIs directly). The mechanical difference is autonomy handling: Operator pauses at every sensitive step to request user confirmation (logins, payments, personal data), requiring the user to stay present. Manus offers its asynchronous cloud model and configurable autonomy modes, so genuinely fire-and-forget execution across parallel tasks is possible. Operator is embedded in ChatGPT's $20/month Plus subscription with no separate credit system; Manus charges its own credit-based pricing on top of a subscription. For research tasks requiring reasoning depth, Operator benefits from GPT's internal reasoning. For breadth research across many sources with minimal supervision, Manus has a structural advantage due to cloud-async parallel execution.

What the agent loop reality looks like

Watching Manus work is genuinely impressive the first few times. You assign a task, it opens a browser window you can observe via real-time screen streaming, clicks methodically through pages, extracts data, and builds a file. The loop feels autonomous in a way that chatbot-driven approaches do not. The problem is what happens at the edges.

Manus fails silently on Cloudflare-protected sites and pages with complex anti-bot measures. Tasks stall mid-execution, consuming credits and returning incomplete results. The CAPTCHA problem is structural: the agent cannot solve modern anti-bot challenges reliably, which limits it on any research task involving news sites, SaaS pricing pages behind paywalls, or corporate directories. Infinite loops are a reported pattern, where the agent cycles through the same steps without progressing, burning credits without output.

"400 credits gone after 4 restaurants on Google Maps. Burned through 1,000 credits before the first generation even finished. $39 for 3,900 credits? You could blow that in one long task." -- aggregated user reports, Reddit and AllAboutAI, 2026

In TechCrunch's March 2025 hands-on test, food ordering crashed after 10 minutes. Flight booking returned only broken links. A Naruto-inspired game build errored out after 30 minutes. Co-founder Alexander Doria acknowledged "error messages and endless loops" during testing. These failure modes are real, even if the platform's overall capability has improved since initial launch.

The credit economics amplify the frustration. A complex research task consumes 500-900 credits. On the Standard plan's 4,000 monthly allotment, that allows 4-8 meaningful runs per month before hitting the ceiling. The 300 daily refresh credits sound useful but deplete in a single moderately complex task. Server load errors ("tasks cannot be created") were frequent during the growth period, though less so post-Meta acquisition.

Who Manus is built for

Manus works best as a delegation tool for repetitive, browser-based knowledge work at scale. Market researchers running competitive pricing sweeps across dozens of sites, technical founders who need an MVP web app scaffolded before hiring engineers, analysts who need information synthesized from many sources into a formatted report: these are the workflows where Manus earns its subscription cost. The asynchronous cloud model means you can assign ten parallel tasks, close your laptop, and return to results. That pattern is genuinely unavailable from chatbots or synchronous agents that require you to supervise each step.

The Extended plan ($200/month, 40,000 credits) makes economic sense for agencies or consultants running Manus as a core production tool across multiple client projects. At that volume, the per-task credit cost becomes reasonable. Startups using it for batch research, batch site generation, or ongoing competitive intelligence have a real use case at this tier.

What Manus is not

Manus is not a coding agent. It can generate simple web apps through browser interaction, but it cannot write, test, debug, and deploy a software project the way Devin or Claude Code can. Do not use it as a substitute for a development environment.

It is not an integration platform. Manus does not connect to Slack, Salesforce, Notion, or most SaaS tools. Workflows that require data moving between business applications need a different tool (n8n, Zapier, or a custom API setup). Manus is a standalone agent that operates in a virtual browser, not a middleware layer.

It is not suitable for tasks requiring login to sensitive personal accounts. Logging in through Manus's cloud virtual browser means session cookies are processed in Manus's (now Meta's) infrastructure. The Team plan includes SOC 2 compliance, but individual users should treat personal banking, primary email, and healthcare portals as off-limits.

The free tier is not a free product; it is a trial. The 1,000 starter credits plus 300 daily refresh credits will sustain basic testing but evaporate during any substantive task. Budget-conscious solo users who need daily automation will hit the ceiling on Standard within a week of heavy use.

Finally: the "Claude wrapper" characterization, while reductive, is grounded. Manus's value is orchestration and UX on top of existing models, not a proprietary reasoning engine. Users who already pay for Claude or GPT-4 are paying for Manus's task-planning and browser-use toolchain, not a fundamentally different underlying intelligence. That may be enough value for the price; it depends on how much time the delegation saves you.