Moonshot AI launched Kimi Hosted Agents on Nov 5, 2026 — managed sandboxes, session memory, and failure recovery as a service. ¥0.072/hour plus tokens, 140+ skills, 8 finance templates.
Moonshot AI officially announced ↗ Kimi Hosted Agents on November 5, 2026, positioning it as "Harness as a Service" — a managed runtime for long-running AI agents that abstracts sandboxes, session memory, and failure recovery. This is distinct from the K3.1 Pro open-weight model release that preceded it; Hosted Agents sells the infrastructure, not the weights.
According to the platform documentation ↗, Moonshot provides:
Isolated sandboxes with network ACL controls, stoppable sessions, and audit export
140+ Skills and 30+ plugins pre-integrated
8 finance Agent templates in the initial official batch, with broader template coverage noted in secondary press
Kimi Inference Hook (enterprise-only): safety check before each inference that inspects inputs and tool returns
Self-hosted sandbox option for enterprise-certified accounts
Per the pricing page ↗:
Sandbox runtime: ¥0.072 per hour
Model inference tokens: billed separately
Plugin call fees: billed separately
The sandbox fee covers the session orchestration layer; token consumption and tool invocations add to the per-session cost. Moonshot's pricing transparency contrasts with competitors that bundle runtime and inference into a single SKU.
Managed runtime framing:
"Moonshot AI launches Kimi Hosted Agents as a managed agent runtime. - @Kimi_Moonshot: sessions, sandboxes, failure recovery as a service - 140+ skills, 30+ plugins; formalizes a Sept. beta - Sandbox ¥0.072/hr plus tokens — Moonshot's own figures"
Launch context alongside open weights:
"Kimi Hosted Agents launched on Nov. 5, renting out the sandboxes, tool calls, and session memory that AI agents need to run. The launch came right after Moonshot put the K3.1 Pro weights on Hugging Face"
Cloud sandbox economics:
"A cloud sandbox for your agent at about a penny an hour? Moonshot just launched Kimi Hosted Agents, the whole harness as a service: sessions, sandboxes, crash recovery, up to 20 sub-agents, 140+ skills."
Teams building long-horizon agents without wanting to operate the harness: Kimi Hosted Agents targets infrastructure buyers, not end-user product agents like Perplexity Teammates. If you need persistent session state, crash recovery, and networked tool execution but do not want to build LangChain-style orchestration in-house, Moonshot's pricing is legible: ¥0.072/hour sandbox + tokens + plugin fees. Compare this with self-hosting agent frameworks or using AWS Lambda + DynamoDB for state.
Finance template early adopters in China: the 8 finance Agent templates and enterprise Inference Hook suggest a compliance-first GTM in regulated verticals. The Inference Hook gates every model call with a safety check — relevant for teams that must audit agent decisions before execution. Self-hosted sandbox for enterprise-certified accounts further signals a buy-vs-build pitch to banks and asset managers.
Kimi K3.1 Pro users considering the stack: K3.1 Pro open weights let you run the model anywhere; Hosted Agents sells the runtime layer on top. If you self-host K3.1 Pro today and hand-roll state + tools, Hosted Agents is the managed alternative. If you only need the model, the weights are free and you can skip the service.
Compared with Perplexity / Claude Code / self-built stacks: Perplexity Teammates ships as an end-user agent product; Kimi Hosted Agents ships as infrastructure. Claude Code is IDE-native and session-scoped; Kimi sessions persist and can resume. LangChain gives you full control and no SLA; Moonshot gives you managed uptime and audit logs at ¥0.072/hour. The buyer persona differs.
International platform availability: platform.kimi.com ↗ is the China-facing domain. platform.kimi.ai ↗ is the international mirror; at publication time, the Hosted Agents documentation may not be fully mirrored there. Verify regional availability if you are outside China.
For deeper context on agent harness options, see our agent frameworks comparison. For open-weight models that could run in a Hosted Agents sandbox, see best open-source LLM 2026. For the Claude MCP protocol that addresses similar tool-calling concerns, see our MCP tutorial.
Kimi: Kimi Hosted Agents announcement ↗ — Nov 5, 2026
AI Affairs — Moonshot AI Kimi Hosted Agents coverage ↗ — Nov 5, 2026