MiniMax Agent is a general-purpose AI agent that plans and executes long-horizon tasks end to end — deep research, decks, full-stack apps and browser work — from a desktop client, the web or your phone.
Used this tool? Rate it
Used this tool? Rate it
MiniMax Agent is a general-purpose AI agent built to carry work through to a finished result rather than answer one question at a time. The vendor's own framing is direct: it is a general intelligent agent designed to tackle long-horizon, complex tasks, launched on 19 June 2025 under the slogan code is cheap, show me the requirement. That slogan is the product thesis in miniature — the scarce thing is a clearly stated requirement, and the agent's job is to turn it into working output.
Mechanically, the product describes itself as more than a single model with tools bolted on. Supported by a multi-agent system, it formulates expert-level solutions through multi-step planning, breaks down task requirements and executes multiple sub-tasks to deliver the final result. In practice that means a request like "research this market and give me a deck" produces a plan, a sequence of sub-tasks and an actual file, rather than a wall of text describing what you should go do.
It comes from MiniMax, a Chinese AI company that builds its own model family, so the agent runs on first-party models rather than reselling someone else's. That vertical integration is the main structural difference from agent products built purely as orchestration layers, and it shows up in the pricing, which is denominated in model tokens rather than opaque credits. On the marketing side the official site sums the pitch up as minimize effort, maximize intelligence.
The product has moved quickly and is worth placing on a timeline before you compare it against reviews you find elsewhere. After the June 2025 launch, the mobile app followed in October 2025 alongside the agent-optimised M2 model, and the desktop app arrived in January 2026 together with a 2.0 web release and Expert Agents. By August 2026 the desktop client is the flagship surface, and much of the newest capability — browser control, long-running goals, persistent memory, a plugin marketplace — lands there first. Anyone reading a 2025 review is reading about a substantially different product.
Subscriptions are sold as token plans rather than seat licences, which is unusual and generally favourable for heavy individual users. Three tiers are offered: Plus at 20 dollars a month, Max at 50 and Ultra at 120, or 200, 500 and 1,200 dollars a year, with the annual option presented as saving roughly two months against the monthly rate.
What separates the tiers is capacity rather than feature access. The site quotes roughly 1.7 billion M3 tokens a month on Plus, 5.1 billion on Max and 12.5 billion on Ultra, with three to four, four to five, and six to seven concurrent agents respectively. Every tier includes the same underlying model family access, a one-million-token context window, native multimodal understanding of image and video input, and a single shared quota across text, image and speech. That shared quota is worth internalising: generating images and speech draws down the same allowance as text, so a media-heavy month costs more than the token figure alone suggests.
Video is the exception to token metering. Hailuo 2.3 video generation is capped at three clips a day on Max and five on Ultra, and is absent from Plus, so anyone whose primary use case is video should treat this as the deciding factor rather than the token allowance.
Beyond subscriptions there is a separate credits option for pay-as-you-go usage, and a daily engagement mechanic: daily check-ins award 400 points, rising to 1,000 on days four and seven for up to 4,000 in a full week, valid for 30 days. Free access exists but is bounded and, by several user accounts, inconsistent from day to day — treat it as a trial mechanism rather than a sustainable free tier.
It is for work that has a defined outcome and too many steps to do by hand: a researched report, a deployed application, a filled-in set of forms, a maintained long-running task. The vendor positions it as a general intelligent agent designed to tackle long-horizon, complex tasks, and that framing describes its sweet spot accurately.
Plus at 20 dollars a month, Max at 50 and Ultra at 120, or 200, 500 and 1,200 dollars a year. The tiers differ mainly by token allowance and concurrency rather than by which features you can reach. A separate credits option and a daily check-in reward scheme also exist.
There is free access, but it is limited and users report it varying from day to day. Daily check-ins award 400 points, rising to 1,000 on days four and seven for up to 4,000 in a full week, valid for 30 days. Treat free usage as a way to evaluate the product rather than as a plan to rely on.
The desktop client ships for macOS on both Apple silicon and Intel, plus 64-bit Windows, alongside a web workspace and an Android app. The desktop client is the most capable surface, and the phone increasingly serves as a remote control for it.
Yes, on the desktop client. It operates a real browser — reading pages, filling forms, uploading files from the current task or workspace, and checking the result after each action — and it can work with local files and the terminal. Outward-facing actions such as publishing or sending ask for confirmation first.
Yes. Proactive Memory retains long-term context such as communication preferences, work habits and project conventions, and applies it in later tasks. You can state your conventions once instead of restating them each session.
Yes, on the paid tiers that include it. Hailuo 2.3 video generation is capped at three clips a day on Max and five on Ultra, and is absent from Plus. Because this is a daily clip cap rather than token metering, video needs should drive your tier choice directly.
Yes, through two routes. The vendor advertises native MCP integration covering GitHub and GitLab, Slack and Figma, and there is a built-in plugin marketplace covering financial data, company lookup, legal search, knowledge workflows and office work, with WPS and WeCom added for office use.
We could not retrieve the product's privacy policy or terms of service during research, so questions about retention periods and training use should be settled against the official documents. What we can report is narrower: the Play listing declares no data shared with third parties, encryption in transit and a route to request deletion; the Chrome cookie import used for browser sessions does not include browsing history, cache, or saved passwords; and in the feedback flow, log upload can be turned off if privacy is a concern.
If software development in a repository is your entire workload, a coding-first tool will likely feel sharper. This product's argument is breadth at a token price: the same subscription that reviews your code also produces the deck, the research and the video. Whether that consolidation is worth a small loss of depth depends on how varied your week actually is.
