Moonshot Sells the Agent Runtime Itself With Kimi Hosted Agents
Moonshot AI has launched Kimi Hosted Agents, a service that packages the agent runtime as a managed product rather than leaving enterprises to build and maintain that layer themselves, and the company is describing it as harness as a service. The phrase is clumsy, but the offer underneath it is clear enough.
The pitch is about the unglamorous part of agents. Teams configure their agents on Kimi's infrastructure and get a hosted environment that runs around the clock for long tasks and for several agents working together, with isolated sandboxes, network access controls, session boundaries an operator can stop at any time and execution logs that can be exported for audit.
But the billing is where the business model shows. Usage reports break out sandbox hours, model inference and plugin calls, and Chinese tech outlet IT Home cites a sandbox rate of 0.072 yuan an hour on top of model and plugin charges, which turns an agent into a metered workload rather than a seat licence. That pricing rewards teams that write efficient agents and punishes the ones that leave a sandbox running all night, which is a discipline most enterprise software does not impose.
The catalogue is built for office work rather than for demos, and the list is the part that decides whether a team can ship without writing its own integrations. The platform lists more than 140 skills and more than 30 plugins covering web search, professional data queries, office documents and content generation, and developers can attach their own tools, data sources and MCP services.
And the control that enterprises will ask about first is the inference hook. An optional component called Kimi Inference Hook lets a company's own or a third party safety service inspect task inputs and tool returns before each model call, so only content that passes the rules reaches inference, and it is limited for now to enterprise verified customers.
And Moonshot is also offering a self hosted option, where sandbox tools run on machines the customer controls while orchestration stays on the platform, which is the arrangement most large buyers want before they point an agent at internal systems.
The templates show where the demand is. IT Home highlights eight finance templates covering modelling, deep research, morning briefs, IPO tracking and portfolio monitoring, while the documentation lists a first wave of 12 that adds game operations and incident response, and each can be copied into a project space and wired to private databases.
So the launch moves Kimi from models and desktop agents into a maintained runtime, and the timing is deliberate. It arrives a day after Huawei Ascend added day zero support for the open weight Kimi K3.1 Pro model, which means the same week brought both cheaper ways to serve Kimi and a managed place to run it. For a Chinese AI company selling into enterprises, that pairing is the argument, since the model, the hardware it runs on and the runtime around it all now come from suppliers the buyer can name.
Moonshot launched Kimi Hosted Agents, a managed runtime for long running and multi agent work
It provides isolated sandboxes, network controls, stoppable sessions and exportable logs
Usage is metered by sandbox hours, inference and plugin calls, with sandboxes at 0.072 yuan an hour
More than 140 skills and 30 plugins are listed, with support for custom tools and MCP services
An optional inference hook screens inputs and tool returns for enterprise customers
Source: Pandaily
