NConnect
DocsKimi K3 + Claude Code

TUTORIAL

How to run Kimi K3 with Claude Code

Use Claude Code with Moonshot AI's Kimi K3 and its full 1M-token context, served by Nebius Token Factory through NConnect. No Anthropic account needed. Video included.

Kimi K3Claude Code 5 minUpdated

Claude Code speaks the Anthropic Messages API, which Nebius does not serve. NConnect runs a local proxy that translates every request to Nebius chat completions, so Claude Code works unchanged while Kimi K3 does the thinking. Kimi K3 is Moonshot AI's frontier open-weights model for coding and agentic tool use.

Watch the walkthrough

Kimi K3 with Claude Code is the Best Combo! · Watch on YouTube

Prerequisites

  • macOS or Linux.
  • A Nebius Token Factory API key from tokenfactory.nebius.com.
  • No Anthropic account or API key. Your existing Claude login is not used.

Steps

  1. Install NConnect

    This installs the CLI and the nclaude alias, plus Bun if needed.

    curl -fsSL https://nconnect.sh/install.sh | bash

  2. Add your Nebius API key

    nconnect configure

  3. Install Claude Code

    Skip this if claude is already on your PATH. In an interactive terminal, NConnect also offers to run this for you the first time you launch Claude Code.

    npm install -g @anthropic-ai/claude-code

  4. Launch Claude Code on Kimi K3

    nconnect --model moonshotai/Kimi-K3 claude

    Pass a prompt straight through to Claude Code for a one-off run:

    nconnect --model moonshotai/Kimi-K3 claude -p "explain this repo"

What NConnect does on launch

  • Starts (or reuses) the local daemon and points Claude Code at it. Anthropic keys, login tokens and model variables are removed from the session's environment.
  • Fills Claude Code's /model menu with Nebius models, each labelled "via nconnect - not Anthropic". Built-in exploration subagents use Kimi K2.7 Code in the Haiku slot.
  • Sends the 1M-context hint for Kimi K3, so Claude Code uses the full window instead of compacting early.
  • Backs the native web_search tool with Tavily when you have added a Tavily key, and routes pasted images to a vision model.
  • Disables /feedback, which would post your transcript to Anthropic rather than through the proxy.

See what the session cost

Claude Code is proxied, so every turn is metered. The total prints when you exit:

On exit
[nconnect cost] session total: $0.0412 (11,840 in, 1,205 out)

The numbers above are illustrative. Run nconnect usage --last 7d for a breakdown by model and tool.

Switch models later

Pick another model from /model inside Claude Code. NConnect remembers it and uses it as the default the next time you run nclaude.

More tutorials

Kimi K3Hermes AgentHow to run Kimi K3 with Hermes AgentRun Nous Research's Hermes Agent on Moonshot AI's Kimi K3 with a 1M-token context, served by Nebius Token Factory through NConnect. Setup takes about five minutes. 5 minRead tutorial Kimi K3Grok BuildHow to run Kimi K3 with Grok BuildUse xAI's Grok Build terminal harness with Moonshot AI's Kimi K3 on Nebius Token Factory. No xAI API key needed, and your Nebius key never reaches api.x.ai. 5 minRead tutorial GLM 5.3Codex CLIHow to run GLM 5.3 with Codex CLIRun OpenAI's Codex CLI on Z.ai's GLM 5.3, a 1M-context model for coding and tool use, served by Nebius Token Factory through NConnect. No OpenAI account needed. 5 minRead tutorial Qwen 3.5 397BOpenCodeHow to run Qwen 3.5 with OpenCodeRun the OpenCode terminal agent on Alibaba's Qwen 3.5 397B, served by Nebius Token Factory through NConnect, without editing your OpenCode config. 5 minRead tutorial