MaxHermes is a self-evolving cloud AI assistant from MiniMax: it deploys in about ten seconds with no server or API key, distils reusable skills as it completes tasks, and works directly inside WeCom, Feishu or DingTalk.
MaxHermes: a cloud AI assistant that writes its own skills
Most AI assistants have their limits fixed at deployment. What you can do with them depends on which tools the vendor preset. MaxHermes attacks that constraint directly by merging two activities that are usually separate: completing a task and accumulating capability. Each complex task it finishes becomes a reusable skill, ready the next time a similar problem appears.
It is a cloud sandbox Hermes from MiniMax, built on Hermes Agent and driven by the MiniMax M2.7 coding model. Three ideas describe it: it runs in an isolated cloud environment, it evolves its own skills at runtime, and it lives inside the chat apps you already use rather than in a separate window.
What self-evolution actually means
Products like OpenClaw depend on skills that a human configures in advance, so the capability set is fixed and you work within it. MaxHermes generates and iterates its own Skills as it works. That single change alters the value curve: instead of peaking on day one, the assistant's coverage widens with use.
Core capabilities
- Autonomous skill evolution: every completed complex task is distilled into a reusable skill, expanding the capability boundary over time.
- Always-on cloud runtime: deployment completes within 10 seconds and the agent stays online 24/7, unaffected by whether your own machine is switched on.
- Direct IM conversation: integrated with WeCom, Feishu and DingTalk, so instructions are issued in the chat window you already have open.
- Persistent cross-session memory: feedback is retained long term and understanding sharpens across conversations.
- Natural-language scheduled tasks: periodic jobs are defined in plain language, such as a daily report or a timed data check.
- Parallel sub-agents: complex work is split across several sub-agents running at once to cut waiting time.
- Token Plan offsetting: task consumption is covered by Token Plan, while tool calls deduct account credits.
Getting started
- Open the MiniMax Agent site and find the MaxHermes entry.
- Click start; no server or API key is required.
- Choose the channel: WeCom, Feishu or DingTalk.
- Converse in that chat app and issue task instructions directly.
- Simply keep using it; skills accumulate and refine as you go.
How it differs from preset-skill agents
Deployment: pure cloud SaaS starting in ten seconds, against local or self-hosted setup. Skills: generated and iterated by the agent, against human-preset and fixed. Barrier: no API key needed, against supplying your own server and credentials. Memory: persistent across sessions, against session-level. Access: inside WeCom, Feishu and DingTalk, against a standalone app or web page. Cost: Token Plan offsetting, against self-managed API billing.
Who it suits
Individuals who want a long-running assistant rather than an occasional chat window; teams automating repetitive office work such as report generation, data tidying and meeting scheduling; developers needing help with code, debugging and documentation; and research roles that want scattered know-how consolidated into reusable capability. The scheduled-task feature is especially practical for operations, where a daily report or a periodic data sweep only needs describing once.
Points of caution
- Self-generated skills deserve review. A skill distilled from one task inherits that task's biases, so high-frequency critical skills should be checked against real results rather than assumed to improve indefinitely.
- Zero barrier is not zero cost. Servers and API keys are removed from the equation, but Token Plan and credits are still consumed, and tool calls are billed separately.
- IM integration is both advantage and constraint. Working inside WeCom, Feishu or DingTalk is convenient, yet the experience depends on each platform's bot and permission model, so cross-platform consistency needs testing.
- The ceiling remains the model. MiniMax M2.7 is strong for coding tasks; for multimodal input or specialised domain knowledge, judge by tests in your own scenarios.
In one sentence
The interesting part is not another assistant but the relocation of skill accumulation from human configuration into runtime behaviour. If that holds, the value curve shifts from peaking at deployment to improving with use.
Frequently asked questions
What is the main difference from an ordinary AI assistant?
Where the skills come from. An ordinary assistant is fixed at deployment by whichever tools were preset, while MaxHermes distils reusable skills from each complex task and iterates them, so coverage widens the longer it is used.
Do I need a server or an API key?
No. It is pure cloud SaaS, with deployment stated at under ten seconds and 24/7 uptime afterwards. You open the MaxHermes entry on the MiniMax Agent site, click start, and pick WeCom, Feishu or DingTalk as the channel.
How is billing calculated?
Task consumption is offset by Token Plan, while tool calls deduct account credits. What you avoid is the upfront server and API key investment, though Token Plan and credits are still consumed, so estimate your usage rhythm before heavy reliance.
Are the automatically generated skills reliable?
They are built from the experience of specific tasks, so a flawed task leaves a flawed skill. For critical high-frequency skills, review the actual outcomes periodically instead of assuming they keep getting better.
Which platforms are supported and can it run scheduled tasks?
It integrates with WeCom, Feishu and DingTalk so you can converse in the chat window directly, and it accepts periodic tasks described in natural language, such as daily report generation or timed data checks.
