Executive summary
On November 5, 2026, Moonshot AI released Kimi Hosted Agents, packaging the harness that runs AI agents into a managed cloud service. Enterprises get 24/7 long-horizon execution, multi-agent collaboration, isolated cloud sandboxes, network access policies, a credential vault and audit export without building their own runtime. Eight financial agent templates ship at launch.
What happened
Moonshot AI announced Kimi Hosted Agents on November 5, 2026 through its official channels. Unlike selling model inference by the token, the product sells the environment that keeps an agent running: the platform owns session state, sandbox lifecycle, tool and plugin calls, and file artifact persistence, while the enterprise configures the agent itself.
The official documentation describes Hosted Agents as a "prebuilt, configurable hosted agent runtime platform." Developers can connect through an API, a console or the kha CLI. The API is served at api.moonshot.cn with an interface version marked 2026-09-01-beta, which signals that quotas and interfaces may still change.
The launch fills a gap in Kimi's line-up. The company already had model capability (K3, K3.1 Pro) and end-user products (the Kimi app, Kimi Code), but enterprises that wanted agents running around the clock still had to assemble scheduling, sandboxing, credential management and auditing themselves. Hosted Agents turns that engineering into a platform capability.
Key facts
| Item | Detail |
|---|---|
| Release date | November 5, 2026 |
| Publisher | Moonshot AI |
| Product | Kimi Hosted Agents |
| Form factor | Managed cloud agent runtime (agent harness as a service) |
| Runtime | 24/7 hosting, long-horizon tasks, multi-agent collaboration |
| Core concepts | Agent (model + system prompt + tools + MCP + skills + plugins), skill, plugin, session, environment, vault |
| Isolation | Dedicated cloud sandbox per session, preinstalled dependencies, init scripts, network policy |
| Security and governance | Network access control, session abort, managed credential vault, audit log export |
| State | Session history, sandbox state and artifacts persisted by the platform; resumable after disconnects |
| Extensibility | 140+ skills and 30+ plugins under the vendor's published figures |
| Access | API, console, kha CLI |
| Pricing inputs | Model tokens + sandbox runtime + plugin calls |
| Launch templates | Eight financial agent templates covering modeling, deep research, morning notes, IPO tracking, position monitoring and related tasks |
From selling models to selling a runtime
For the past two years, model vendors mostly sold inference by the token. Agents change the cost structure: a single task may call the model repeatedly, read and write large contexts, invoke tools and plugins, execute code in a sandbox and run for hours or days across many turns.
Under that load, the hard problem is not whether the model can answer but whether the task completes reliably. How is the sandbox isolated? How are dependencies installed? How is network access restricted? Where do credentials live? How does a session resume after a disconnect? How are actions audited afterwards? These are harness engineering questions, not model capability questions.
Kimi Hosted Agents productizes that layer. The official concepts cover the full anatomy of an agent: model, system prompt, tools, MCP, skills and plugins define the agent; a session carries multi-turn interaction and state; an environment provides the cloud sandbox and init scripts; a vault centralizes credentials for external systems. The platform orchestrates these pieces and scales them on demand.
The immediate benefit is a shorter path to production. Instead of building an agent runtime, a company can configure an agent in the console, attach skills and plugins, set a network policy and run it. The trade-off is dependence on the vendor for runtime, billing and governance, so migration cost deserves an early assessment.
Platform capabilities: sandbox, state and audit
The published capabilities fall into four layers.
Execution isolation. Each session runs in its own cloud sandbox, with preinstalled dependencies, initialization scripts and a configurable network access policy. For finance and healthcare, that means "which external services this agent may reach" becomes an explicit rule rather than an application-level convention.
State persistence. Session history, sandbox state and file artifacts are stored by the platform. An agent can continue a task across turns, and artifacts survive process restarts. For long-horizon work, this layer decides whether a ten-hour task is feasible at all.
Governance and audit. The platform provides session abort, a managed credential vault and audit log export. Enterprises can keep keys in the vault instead of hard-coding them, then export call records for compliance review. This is usually the first question procurement asks.
Extension ecosystem. Agents can mount skills and plugins, quoted at 140+ skills and 30+ plugins. Skills package reusable task flows; plugins connect to external systems. In finance, the plugin layer determines whether data sources such as Wind, iFinD or Hundsun can be reached.
First use cases: financial agent templates
At launch, Moonshot published eight financial agent templates covering investment modeling, deep research, morning note generation, IPO tracking and position monitoring. Their value is not that they are extra samples but that they decompose recurring research workflows into reusable agent configurations: data sources, analysis steps, output formats and review stages are predefined.
The published investment research assistant uses one routing agent plus 13 specialist sub-agents. Six execution agents cover research analysis, fundamental modeling, backtesting, portfolio risk, market strategy and fund-flow analysis. Seven review agents examine conclusions from value, growth, contrarian, macro, quality and momentum angles, and a final aggregation step merges the output. The agent runs on kimi-k3 and mounts the enterprise edition of a financial investment analysis plugin connected to Wind, iFinD, Hundsun, Caixin, Xinhua Finance and Tianyancha.
Moonshot's efficiency claim is that a single company research memo previously took one to three working days of human effort, while the official assistant is expected to finish in tens of minutes. That figure is vendor-reported; actual latency depends on data entitlements, task complexity and model quota. The company also states that the assistant performs research and analysis only, and does not execute trades.
Beyond finance, Moonshot describes two enterprise directions: internal system integration, such as business monitoring and operations response, and user-facing product features, such as an in-app research assistant or customer service agent. The former emphasizes permissions and audit, the latter concurrency and latency, and they place different demands on the platform.
Pricing, access and limits
Pricing has three components: model tokens, sandbox runtime and plugin calls. That differs from token-only API pricing. A long-running task can be cheap in tokens but expensive in sandbox hours; heavy use of external data sources raises plugin costs. Buyers estimating unit economics need all three inputs rather than a per-million-token comparison alone.
There are three access paths: the API for teams with existing systems, the console for quick validation and the kha CLI for local development and debugging. All three share the same agent configuration, which simplifies the move from test to production.
The limits deserve equal attention. Hosted Agents solves runtime, not outcomes: task decomposition, prompts and tool selection remain the enterprise's design, and the platform does not guarantee that a business conclusion is correct. The financial templates provide a process skeleton; data licensing, compliance review and result verification stay with the customer. The service is also in beta, so interfaces, quotas and pricing may change, and critical workloads should keep a fallback path.
Official sources
- Shanghai Securities News via Sina Finance: Kimi launches Hosted Agents
- Beijing Business Today via Sina Finance: Moonshot introduces hosted agent service
- Kimi Hosted Agents quickstart: Hosted Agents Quickstart
- Kimi official investment research agent: Official Investment Research Agent
- Kimi help center agent overview: Agent Overview