Kimi Hosted Agents Operational

Kimi Hosted Agents is an enterprise-grade hosted agent service launched by Moonshot AI. Positioned as Harness as a Service, it provides the cloud-based execution environment, session management, tool invocation, permission controls, and other frameworks required to run agents. Enterprises can access a 7×24-hour runtime without building one themselves, with support for long-running tasks and multi-agent collaboration. Kimi Hosted Agents includes 140+ skills and 30+ plugins, along with sandbox isolation and sensitive-content filtering. Billing is based on tokens, sandbox duration, and plugin calls.

Cloud-hosted harness: Provides the entire agent runtime framework, including the execution environment, session management, tool invocation, permission controls, interruption recovery, and more, as a cloud service. Enterprises can access a 7×24-hour runtime without building one themselves.
Long-running tasks and multi-agent collaboration: Supports continuously running long tasks and coordinated work among multiple agents, making it suitable for time-consuming scenarios such as business monitoring and deep research.
Sandbox and permission controls: Provides isolated sandboxes and network access controls. Enterprises can define operational boundaries, terminate sessions at any time, trace execution history, and export audit records.
Skills and plugin library: Includes 140+ skills and 30+ plugins covering web search, professional data queries, office document processing, and more. It also supports integrating enterprise-owned tools and data sources.
Kimi Inference Hook: An independent security component that checks inputs and tool outputs before each model inference, blocking privacy and sensitive data according to enterprise-defined rules. It is currently available only to selected enterprise-authenticated users.
Official agent templates: The first release provides 12 templates covering financial research and modeling, deep research, morning reports, operations response, and other scenarios. Templates can be copied into a project space, connected to internal data, and quickly customized and deployed.

Registration and authentication: Visit https://platform.kimi.com/docs/hosted-agents/quickstart to register an Open Platform account, then create an API Key under “Organization Management—API Management.”
Install the kha CLI: Install kha using the official installation script and log in with your API Key. You can then manage agents and sessions using natural language in AI coding assistants such as Kimi Code, Claude Code, and Codex. You can also use the console website or REST API directly.
Select an agent: View the available agent list through the API or console, select the Agent that will handle the task, and save its ID.
Create an execution environment: Configure the cloud sandbox environment where the agent will run.
Create a session: Associate the agent with the execution environment to create a session, which serves as the carrier for task execution and state persistence.
Subscribe to the event stream: Subscribe to the session event stream through SSE to receive the agent’s reasoning, tool calls, and message output in real time. After disconnection, resume reading with Last-Event-ID.
Send a task message: POST a user.message to the session with the specific task. The agent will immediately begin calling tools to execute it.
Wait for completion and retrieve artifacts: When the session status appears as idle in the event stream, the task is complete. Use the Artifact API to list and download delivered files. Intermediate files can also be browsed and downloaded directly from the session file system.

No self-built harness: The underlying frameworks for the execution environment, session management, tool invocation, permission controls, interruption recovery, and more are fully hosted in the cloud. Enterprises avoid months of in-house development and ongoing maintenance, allowing them to focus on business application development.
7×24 long-running execution: Session state and files are persisted on the server, supporting asynchronous long-running tasks and multi-agent collaboration. This is suitable for scenarios requiring continuous operation, such as business monitoring and deep research.
Secure and controllable: Isolated sandboxes combined with network access controls allow enterprises to define operational boundaries, terminate sessions at any time, trace execution history, and export audit records.
Pre-inference security filtering: Kimi Inference Hook checks inputs and tool outputs before each model inference, blocking privacy and sensitive data before it enters the model. Inspection rules are defined by the enterprise.
Ready-to-use capability library: Includes 140+ skills and 30+ plugins covering common tasks such as web search, professional data queries, and office document processing. It also supports integrating enterprise-owned tools and data sources.
Fast deployment with official templates: The first 12 official Agent templates cover financial research and modeling, operations response, and other scenarios. Copy them into a project space, connect internal data, and customize them for deployment, significantly shortening the path from prototype to production.

Project website: https://platform.kimi.com/docs/hosted-agents/overview

Investment research modeling and deep research: Connect financial data plugins and research report data sources to automatically break down tasks and call tools for financial modeling and industry research, continuously producing investment research results 7×24.
Financial morning reports and content generation: Aggregate market data and news on a schedule to automatically generate morning reports, PPTs, and other office documents, supporting content production with fixed schedules such as institutional morning meetings.
IPO tracking and portfolio monitoring: Continuously track Hong Kong stock IPOs and portfolio holdings on a scheduled basis, triggering alerts when anomalies occur and serving as a tireless monitoring agent.
Enterprise operations and business monitoring: Integrate with internal systems to execute operations response and business monitoring tasks, automatically handling alerts and routine inspections within sandbox and permission boundaries.