Free Lesson
Run Claude Code on Kimi K3 and Build a Research Agent
60 min
Sep 12, 2026 11:00 AM
By continuing, you agree to Maven's Terms and Privacy Policy.
What you'll learn
Point Claude Code at Kimi K3, one project or all if you want
Nine environment variables in one settings file. Every other project you own stays on Claude, untouched.
Build a research agent live and run it on your question
Subagent, MCP server, slash command. The room picks the research question and we read the result together.
Read the real bill and decide what to move
Measured cost per run, not a pricing page. You leave having named one project you will move and one you never will.
Why this topic matters
Claude Code does not have to run on Claude. You can point one project at another model, keep the rest on Claude, and pay a fraction on work that does not need frontier reasoning. The question is which work. I ran the same research task on three engines. $0.3183 on Claude Opus 5, $0.2485 on Kimi K3, $0.0049 on DeepSeek. Kimi was also the slowest, 400 seconds against 210.
