Moonshot AI has formally launched Kimi Hosted Agents, a managed cloud service that runs enterprise AI agents around the clock without asking customers to build or maintain the agent framework themselves. The service, announced on 5 November, packages sessions, tool execution, sandboxes and failure recovery into a hosted runtime on the Kimi API Open Platform. It supports long-running and asynchronous work, lets several agents collaborate, and is aimed first at domestic enterprise customers in China. Pricing is pay-as-you-go, starting with sandbox time at 0.072 yuan per hour.
What the service does
Hosted Agents turns the layer around the model into a product. The harness manages session history, runs built-in tools inside isolated cloud sandboxes, handles retries and scheduling, and resumes work after a platform failure. Developers reach it through a web console, a programming interface, or the kha command-line tool, which works inside Kimi Code, Claude Code and Codex.
The platform ships with more than 140 skills and more than 30 plugins covering web retrieval, data queries, office documents and content generation. Twelve starter templates are included, eight of them for finance work. The launch extends the product line that began with Kimi K3.1 Pro, moving Moonshot from supplying a model to operating the environment around it.
Moonshot pitches the service at internal business monitoring and operations response, as well as customer-facing investment research and customer-service assistants. Long-running work is the point: agents that watch, investigate and report over hours, rather than chatbots that answer one question and stop.
Safety hooks and self-hosting
Enterprises get a safety checkpoint called the Kimi Inference Hook, which can connect to a company’s own security service and inspect task inputs and tool returns before each model call. The hook is offered to enterprise-verified users. Hosted Agents also provides isolated sandboxes, network-access controls and execution logs that can be exported for audit, with sessions operators can stop.
A self-hosted option moves sandbox tool execution onto customer machines, while orchestration, model calls and session management stay on Moonshot’s platform. Moonshot is explicit that this is not private deployment, because tool inputs and results still pass through its model interface. The self-hosted version drops tools, including IPython and image search, and uploaded files are not mounted on the host.
Pricing and availability
Pricing is usage-based across three parts: sandbox runtime at 0.072 yuan per hour, model inference at the selected model’s token rates, and plugin charges by call count and unit price. At roughly one US cent per sandbox hour, the runtime itself is nearly free; the model tokens and plugin calls are where the bill will land.
The service follows a beta that opened in September for domestic enterprise-verified users, so the November announcement formalizes a tested product rather than launching an unproven one. Documentation sits on Moonshot’s Chinese open platform, and the international platform carried no Hosted Agents pages at the time of checking, so the launch reads as China-first for now.
Access is aimed at enterprises and developers, with the Inference Hook limited to enterprise-verified users. No consumer version has been announced, and there is no sign of a free tier. For now the buyer is a team with long-running agent work and a security review to satisfy, not an individual looking for a personal assistant.
Our read
Hosted Agents is Moonshot selling the harness, not the model. Every lab now has a capable model; the scarce piece is a runtime that keeps agents working for hours, recovers from failures and proves what happened. Finance templates point to the first buyers. The race is running in Europe, where Mistral Studio made agents its pitch the same day.
The honest limit is geography. A China-first launch, documented on the Chinese platform with no international pages yet, means most Western teams cannot simply sign up. The Inference Hook and audit logs also suggest Moonshot knows its buyers are regulated firms, not hobbyists.
If the service reaches the international platform with the same terms, it becomes a genuine rival to the hosted-agent offerings from Western labs. Until then it is a strong product in a bounded market, and its real test is whether enterprises trust a Chinese-hosted runtime with their internal systems.
Why It Matters
Agent runtimes are becoming the new platform war. Models are converging in capability, so the contest moves to who can run agents reliably for hours, keep them inside their permissions and show an auditor what happened. Hosted Agents gives Moonshot a claim in that contest, and pressures rivals to match its sandboxes, skills and safety hooks in one managed service.
Background
Moonshot, founded in Beijing in 2023, built Kimi into a leading Chinese AI assistant known for long context. Kimi K3.1 Pro followed in October with a million-token window, and open weights followed in November. The lab has also faced questions about training practices, with reporting covered on this site accusing it and DeepSeek of learning from OpenAI’s Astra outputs.
Source: TechInAsia’s report and Unite.AI’s coverage
0 Comments