Kimi + OpenClaw - Open-Source AI Agent Platform

Model snapshot is historical (verified 2026-10-02)

The note’s original model (Kimi K2.5) was retired across Moonshot’s platform on 2026-08-31 (per Moonshot’s Kimi-in-OpenClaw guide; the pricing page lists it as retired); Kimi Claw now configures Kimi K3 automatically (kimi-k3). K3 has a 1,048,576-token context; K2.6 and K2.7 Code have a 262k context. Current API prices: see Moonshot’s pricing page https://platform.kimi.ai/docs/pricing/chat. The earlier “9x cheaper than Claude” framing no longer holds. For Anthropic pricing see anthropic-claude.

What this is

Two related things share this note:

1. openclaw: the MIT-licensed (per its LICENSE file) self-hosted agent runtime (OpenClaw Foundation, an independent 501(c)(3); 391k GitHub stars on 2026-10-02 per the GitHub API; Node 24.16+ or 26.1+; local Gateway on 127.0.0.1:18789 by default; Markdown memory files SOUL.md, USER.md, IDENTITY.md, TOOLS.md, HEARTBEAT.md; 20+ messaging channels). Full details, install and security history (CVE-2026-25253) live in the openclaw note.
2. Kimi Claw: Moonshot AI’s hosted OpenClaw on kimi.com: one-click cloud deployment, no VPS, 5,000+ ClawHub skills, 40 GB cloud storage, K3 configured automatically; requires an Allegretto membership or higher, while linking an existing local OpenClaw is free (Kimi’s Kimi Claw page, updated 2026-08-12, fetched 2026-10-02). Launched in beta mid-Feb 2026 (MarkTechPost, 2026-02-15). Prices are not on that page. Hosted-OpenClaw peer that was discontinued: kiloclaw.

Kimi K2.5 and Agent Swarm (Jan 2026, now retired)

• 1T-parameter Mixture-of-Experts, 32B active, 384 experts (8 per token), 256K context, native multimodal (400M MoonViT encoder), Modified MIT license; Hugging Face card lists release 2026-01-29.
• Agent Swarm: the model self-directs up to 100 sub-agents and up to 1,500 coordinated tool calls, trained with Parallel-Agent RL; up to 4.5x lower wall-clock time than a single agent (Moonshot tech report, arXiv 2602.02276).
• Model card benchmarks: BrowseComp 78.4% in swarm mode vs 60.6% baseline; WideSearch 79.0% vs 72.7%.
• Video-to-code, screenshot-to-website and visual debugging were headline multimodal features.

Architecture (OpenClaw)

Gateway routes channels to agent sessions; the agent runtime resolves the session, assembles context from memory, streams the model response while running tools, and persists state to disk. Direct-message sessions are sandboxed by default; allowlists control who can talk to the agent. Multi-agent routing lets one instance map different channels to different agents and models.

Security

OpenClaw needs deep system access (files, shell, APIs). Run it on an isolated machine or VPS, keep the gateway off the public internet (SSH tunnel or tailnet), and keep it updated; see openclaw for the advisory list. Using Moonshot’s API sends prompts to Moonshot’s servers; self-hosting OpenClaw with a local model keeps data local.

Related

kimi-k2-dot-5, kimi-k3, kimi-claw-agent-zero-comparison, openclaw-vs-agent-zero, openclaw-nanoclaw-picoclaw-comparison, model-context-protocol

Limits

The Kimi Claw price is not stated on Kimi’s page. The 2026-02 cost analysis and install walkthrough were removed as dated; see openclaw.

Sources

• https://huggingface.co/moonshotai/Kimi-K2.5 (2026-10-02)
• https://arxiv.org/html/2602.02276v1 (Kimi K2.5 tech report, via search; 2026-10-02)
• https://github.com/openclaw/openclaw (2026-10-02)
• https://platform.kimi.ai/docs/pricing/chat and https://platform.kimi.ai/docs/guide/use-kimi-in-openclaw (2026-10-02)
• https://www.kimi.ai/resources/kimi-claw-introduction (2026-10-02)
• https://api.github.com/repos/openclaw/openclaw (2026-10-02)
• https://www.marktechpost.com/2026/02/15/moonshot-ai-launches-kimi-claw-native-openclaw-on-kimi-com-with-5000-community-skills-and-40gb-cloud-storage-now/ (secondary)