Schedule a Call
To set up Kimi K3, get an API key from OpenRouter (model ID moonshotai/kimi-k3) or from Moonshot’s Kimi platform (model ID kimi-k3), then paste the base URL, key and model ID into your coding tool’s settings. For Claude Code, you point it at Moonshot’s Anthropic-compatible endpoint instead, and the exact settings are below. Kimi K3 from Moonshot AI went open source this summer, and everyone’s been talking about it since.
But then you go looking for “how do I actually use this thing” and every guide only covers one tool. One post is just for Cursor. Another is just for Codex.
None of them show you the whole picture. Okay so this post fixes that — one guide, every harness, and you’ll be up and running in under 5 minutes. No coding background needed, honestly, it’s mostly copy and paste.
What Is Kimi K3, and Why Is Everyone Talking About It?
Kimi K3 is a new AI model from Moonshot AI, a Chinese AI lab, and it just went open source. That means the “weights” — basically the trained brain of the model are public. Anyone can download it, run it on their own servers, or plug it into a coding tool.
Think of it like the difference between a restaurant dish and a recipe. A closed-source model like Claude or GPT-4o is the dish — you order it, you eat it, you never see how it’s made. Kimi K3 is the recipe. Anyone can pull it apart, tweak it, host it themselves, or build a business on top of it.
And that’s a big deal. According to Bloomberg’s coverage, Kimi K3 is the largest open-weight model publicly available, with 2.8 trillion parameters, roughly double the size of DeepSeek’s V4 Pro model. Within days of release, 20+ inference providers were already offering it.
At BULDRR AI, we ran our own comparison of Kimi K3 against Claude across coding, legal document review, and everyday task cost, and here’s what stood out:
⚠️ A quick honesty note: these are figures from our own testing and comparison, not third-party audited benchmarks — treat them as a starting point, not gospel, and always run your own comparison for your specific use case before switching your whole workflow over.
One thing that IS independently confirmed: Kimi K3 outperforms Claude’s prior-generation models on coding and agent tasks, and developers have been picking it over closed-source competitors specifically for front-end work.
Get Your API Access (2 Options)
Before you touch any harness, you need an API key. There are two paths — pick whichever fits how you want to work.
Option A — OpenRouter (Recommended)
OpenRouter (new to it? start here) is a router that gives you one API key for Kimi K3 and 200+ other models. Every single harness in this guide supports it, which makes it the easiest starting point — especially if you might want to switch models later.
Sign up at openrouter.ai → Keys → Create Key. Kimi K3 is a paid model on OpenRouter (no :free version as of October 2026): about $0.67 per million input tokens and $14 per million output tokens, with a 1M-token context window, so add a few dollars of credit first. Then you’ll have:
Base URL : https://openrouter.ai/api/v1
Model ID : moonshotai/kimi-k3
API Key : your-openrouter-key
Option B — Moonshot Platform (Direct)
Use this if you want to build directly on Moonshot’s own platform, need Kimi-specific features, or want to run Kimi K3 inside Claude Code.
Sign up at platform.kimi.ai → API Keys → Create Key
Base URL : https://api.moonshot.ai/v1
Model ID : kimi-k3
API Key : your-moonshot-key
💡 Tip: providers update their base URLs and model IDs fairly often as new versions ship. If a direct connection doesn’t work right away, double check the exact endpoint on Moonshot’s current docs before assuming something’s broken on your end.
How to Set Up Kimi K3 in Each Harness
Pick your tool below and copy the config. That’s really it.
Claude Code
Claude Code talks to Anthropic-style APIs, so use Moonshot’s Anthropic-compatible endpoint with your Moonshot key (Option B). Per Moonshot’s official Claude Code guide, add this to ~/.claude/settings.json:
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.moonshot.ai/anthropic",
"ANTHROPIC_AUTH_TOKEN": "your-moonshot-key",
"ANTHROPIC_MODEL": "kimi-k3[1m]",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "kimi-k3[1m]",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "kimi-k3[1m]",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "kimi-k2.7-code",
"ANTHROPIC_DEFAULT_FABLE_MODEL": "kimi-k3[1m]",
"CLAUDE_CODE_SUBAGENT_MODEL": "kimi-k3[1m]"
}
}
✅ Restart Claude Code and run the verification prompt below. Two gotchas from Moonshot’s guide: set every model variable (leaving one out makes the matching tasks fail), and the env block in settings.json overrides variables with the same name exported in your terminal. OpenRouter can also feed Claude Code, but its own docs warn that Claude Code is optimized for Anthropic models and may not work correctly with other providers, so the Moonshot endpoint is the safer route for Kimi K3. New to Claude Code itself? Our guide to using Claude Code for free covers the basics.
Codex CLI
Add this to your ~/.zshrc or ~/.bashrc for a permanent setup:
export OPENAI_BASE_URL="https://openrouter.ai/api/v1"
export OPENAI_API_KEY="your-openrouter-key"
codex --model "moonshotai/kimi-k3"
✅ Once the two exports are saved, run the codex command on the last line and you’re using Kimi K3.
Pi Agent
Open ~/.pi/config.yaml and add:
model: moonshotai/kimi-k3
base_url: https://openrouter.ai/api/v1
api_key: your-openrouter-key
effort: high
Aider
Run this directly from the command line:
aider \
--model openrouter/moonshotai/kimi-k3 \
--openai-api-key your-openrouter-key \
--openai-api-base https://openrouter.ai/api/v1
Or save it permanently in ~/.aider.conf.yml:
model: openrouter/moonshotai/kimi-k3
openai-api-key: your-openrouter-key
openai-api-base: https://openrouter.ai/api/v1
Cursor
Go to Settings → Cursor Settings → Models → Add Custom Model and enter:
Model name : moonshotai/kimi-k3
Base URL : https://openrouter.ai/api/v1
API Key : your-openrouter-key
✅ Save it, select the model from your dropdown, and you’re set.
Any Other OpenAI-Compatible Tool
Wait — what if your tool isn’t listed above? No worries. These three values work for basically any harness that accepts OpenAI-compatible settings:
BASE_URL : https://openrouter.ai/api/v1
API_KEY : your-openrouter-key
MODEL : moonshotai/kimi-k3
Just drop them into whatever config screen your tool uses, and it’ll connect the same way.
Verify It’s Actually Working
Run this exact prompt in your harness of choice:
“Who are you and what model are you running on?”
Expected answer: it should identify itself as Kimi, made by Moonshot AI.
If it says Claude or GPT instead, your setup didn’t take — go back and double check your BASE_URL and MODEL settings. That’s the #1 cause, by the way, of “it’s not working” reports.
The Real Way to Save Money — Planner + Worker Setup
Here’s the part most setup guides skip entirely: don’t use Kimi K3 for everything. Use it only as the planner, and hand the actual grunt work to a cheaper, faster model.
Think of it like a construction site. You don’t need your most expensive architect laying bricks all day — you want them drawing up the blueprint. Then a crew of workers executes the plan. Kimi K3 is your architect. A fast model like Gemini Flash, GPT-4o-mini, or Kimi K2.7 Code is your crew.
Before: One expensive model handles every single step of a project — planning AND every tiny file edit — burning through your budget fast.
After: Kimi K3 plans the architecture once, then a cheap fast model executes the repetitive parts. Same output quality, dramatically lower spend.
In our own comparison of this planner/worker split versus using one powerful model for everything, we saw the combined cost come in at roughly 9 to 10x lower — same result, a fraction of the bill. Again, that’s our own testing, so run a small test on your own project before betting your whole budget on the number.
Common Questions People Ask
Is Kimi K3 really free to use?
The model weights themselves are free and open source — you can download and self-host them at no cost. But if you’re using it through OpenRouter or Moonshot’s hosted API (the easiest path for most people), you pay per token, same as any other AI API. On OpenRouter that’s about $0.67 per million input tokens and $14 per million output tokens as of October 2026.
Do I need to know how to code to set this up?
No. Every setup in this guide is copy-paste — you’re pasting a few lines into a settings screen or a config file, not writing new code. If you can copy a line of text, you can do this.
What if it says “Claude” or “GPT” instead of “Kimi” after setup?
That means your base URL or model ID is wrong somewhere. Go back to the harness section above, re-copy the exact values, and try the verification prompt again.
Which harness should a beginner start with?
If you’re not sure, start with OpenRouter — it works with every tool in this guide, and you only need one key to try all of them. If you live in Claude Code, use the Moonshot key and the Claude Code settings above instead.
Can I use Kimi K3 in Claude Code?
Yes. Moonshot publishes an Anthropic-compatible endpoint (https://api.moonshot.ai/anthropic) for exactly this. Put the env block from the Claude Code section into ~/.claude/settings.json with your Moonshot API key, restart Claude Code, and run the verification prompt.
What is the Kimi K3 model ID?
It’s moonshotai/kimi-k3 on OpenRouter and kimi-k3 on Moonshot’s own API, written kimi-k3[1m] in Moonshot’s Claude Code settings.
Quick Recap
- Kimi K3 is Moonshot AI’s new open-source model — the weights are public and free to download
- OpenRouter (moonshotai/kimi-k3) is the easiest way to connect it to any OpenAI-compatible harness with one key
- Claude Code connects through Moonshot’s Anthropic-compatible endpoint, https://api.moonshot.ai/anthropic
- Codex CLI, Pi Agent, Aider, and Cursor all connect with the same basic base URL / model ID pattern
- Run the verification prompt to confirm your setup actually worked
- Use Kimi K3 as your “planner” and a cheaper model as your “worker” to cut costs significantly
What Should You Do Next?
Your one action right now: grab a key at openrouter.ai (or platform.kimi.ai for Claude Code), add a few dollars of credit, paste it into whichever harness you already use, and run the verification prompt above. That’s it. Nothing to install and nothing to break, and a handful of test prompts costs pennies at Kimi K3’s per-token price.
If you’re curious about pairing Kimi K3 with automated workflows instead of just coding tools, check out BULDRR AI’s beginner’s guide to n8n — it’s a good next step once you’ve got Kimi K3 wired up. And if you want ready-made automations instead of building from scratch, BULDRR AI’s free workflows library has over 10,000 templates you can just download and use.
Each button opens the assistant with the prompt already written, asking it to teach this article step by step, in simple words.
ChatGPT
Claude
Grok
Perplexity
Add as a preferred source
