Kimi Hosted Agents

Kimi Hosted Agents is Moonshot AI's Harness-as-a-Service product, launched on the Kimi Open Platform on 2026-11-05. It packages the sandboxes, sessions, tool execution and failure recovery that agents need as a managed cloud runtime, so teams get a 24/7 environment for long-horizon and multi-agent work without building the harness themselves.

It landed around the same time as open weights for Kimi K3.1 Pro. Teams can embed Hosted Agents into internal systems or customer-facing products via console, kha CLI and REST API, with starter templates (including finance).

What problem it solves

Production agents usually need months of harness work: sandboxes, session state, tools, permissions, recovery and audit. Hosted Agents moves that infrastructure to the platform so teams configure business agents instead of maintaining runtimes.

Core capabilities

• Isolated cloud sandboxes with network controls and stoppable sessions; auditable traces.
• 140+ Skills and 30+ plugins, plus custom skills/plugins/MCP/tools.
• Optional Kimi Inference Hook (enterprise) to gate content before each model call.
• Verified enterprises can run sandbox tools on self-hosted machines.

Pricing

Roughly sandbox-hours × ¥0.072 + model tokens + per-plugin call fees.

Limits

• Tied to the Kimi platform and model catalog; full self-host needs other stacks.
• Inference Hook / self-hosted environments are enterprise-gated.
• Positioned for enterprise hosting, not a personal “cloud teammate” product.

Who it's for

• Enterprises that want long-running/async agents without building sandboxes;
• Teams already on Kimi model APIs that need a managed runtime.

Getting started

1. Read the overview and pricing docs.
2. Create an API key and install the kha CLI (curl -fsSL https://ha-static.moonshot.cn/kha/install.sh | sh).
3. Create an Agent, Environment and Session via console or CLI, attach Skills/plugins, and run a first task.
4. Watch sandbox and token usage reports before rolling out to production.