Moonshot AI has launched Kimi Hosted Agents, a managed cloud service for running AI agents, Pandaily reported on 5 November. The Beijing-based company is offering enterprises a hosted runtime with isolated sandboxes, session controls and a safety check before model calls, rather than requiring them to operate that layer themselves.
Key points
- Hosted Agents provides isolated sandboxes, network-access controls and execution logs that can be exported for audit.
- Kimi Inference Hook lets an enterprise safety service check inputs and tool returns before each model call.
- The platform lists more than 140 skills and more than 30 plugins, alongside agent templates.
Moonshot takes on the agent runtime
Moonshot operates the Kimi chatbot and released Kimi K3.1 Pro, a model with about 3 trillion parameters and a 1 million token context window, Caixin Global reported. K3.1 Pro took fourth place on an AI index. Hosted Agents extends the product line from supplying a model to operating the environment in which an agent uses one.
An agent runtime keeps a task going between model calls: it passes work to tools, brings their results back and manages the session in which those steps occur. Moonshot packages that layer as a hosted service for long-running tasks and collaboration between agents. Enterprises can configure agents on Kimi’s infrastructure instead of building and maintaining the orchestration layer themselves.
The distinction matters when an agent needs to do more than answer a single request. Each tool call has to run somewhere, and the work needs boundaries that an operator can control. Kimi Hosted Agents uses isolated sandboxes with network-access controls, separates work into sessions that operators can stop at any time, and keeps execution logs that can be exported for audit.
Usage reports separate sandbox time, model inference and plugin calls. IT Home puts the sandbox rate at 0.072 yuan per hour, with model and plugin charges added separately, Pandaily reported. Those are distinct parts of the service: a sandbox provides a place for work to run, while model calls and plugins supply responses and access to other functions.
Kimi Inference Hook checks each model call
Kimi Inference Hook adds an optional inspection point before each call to the model. An enterprise can connect its own safety service, or one supplied by a third party, to check both the task input and material returned by tools. Content proceeds to inference only if it passes the enterprise’s rules. The hook is currently limited to enterprise-verified customers.
That placement gives the check a specific job. A tool might return text gathered during a task, and the hook lets the connected safety service inspect that text before it becomes part of the next model call. It also checks the task input before that call. The inspection point sits in the repeated cycle of an agent’s work, rather than only at the start of a session.
Enterprises can point agents at internal systems for monitoring, research or operations work, or put them in customer-facing products as research and support assistants. In either setting, the hook offers a place to apply the enterprise’s rules to material entering the model.
Kimi lists 12 templates and a self-hosted option
More than 30 plugins and more than 140 skills are listed for the platform, spanning content generation, office documents, professional data queries and web search. Developers can attach their own tools and data sources, as well as MCP services, which provide a standard way for an agent to connect to external tools. That gives teams a choice between the listed functions and connections to systems they already use.
IT Home highlighted eight finance templates covering modelling, deep research, morning briefs, IPO tracking and portfolio monitoring. Moonshot’s documentation lists a broader initial set of 12 that also includes game operations and site reliability engineering incident response, Pandaily reported. Templates can be copied into a project space and connected to private databases or internal reports.
Preparing a morning brief from internal reports could use one of those templates if the reports were connected to it. The work could run inside a session that can be stopped, with an execution log available for export afterwards. Hosting the runtime would also remove the need to operate that part of the system locally.
Moonshot’s documentation also describes a self-hosted arrangement in which sandbox tools run on machines controlled by the customer while orchestration remains on the platform.
