Run Claude Code for Free With Kimi K2.5 (5 Steps)

free-claude-code is an open-source proxy with 27k+ stars that sits between the Claude Code client and any model provider you choose. You get the full Claude Code experience, slash commands, tool use, streaming, the works, without paying Anthropic directly.

This guide walks through setting it up with Kimi K2.5 as the model backend. Five steps, about three minutes.

What this actually is

A FastAPI proxy server that exposes Anthropic-compatible endpoints. Claude Code thinks it's talking to Anthropic's API. The proxy intercepts the request and routes it to whatever model provider you've configured.

11 providers supported out of the box: Kimi, DeepSeek, OpenRouter, NVIDIA NIM, Ollama, LM Studio, llama.cpp, Wafer, OpenCode, Z.ai, and more. You can route different model tiers (Opus, Sonnet, Haiku) to different providers.

Kimi is a strong choice because it implements the Anthropic Messages protocol natively. No format translation between OpenAI and Anthropic message formats. Cleaner connection, lower latency.

Prerequisites

Node.js (for Claude Code CLI): npm install -g @anthropic-ai/claude-code
Python 3.14
uv (Python package manager): docs.astral.sh/uv

On macOS, if you don't have uv: brew install uv. On Windows, use the PowerShell installer from the uv docs.

Step 1: install the proxy

One command:

uv tool install --force git+https://github.com/Alishahryar1/free-claude-code.git

This pulls the repo and installs it as a CLI tool. You get two commands: fcc-server (the proxy) and fcc-claude (the modified CLI).

Step 2: start the server

fcc-server

The server starts on localhost:8082by default. You'll see a line pointing you to the admin dashboard.

Step 3: configure Kimi K2.5 as your model

Open the admin dashboard in your browser (the URL the server printed). Navigate to Model Config.

Set the default model to:

kimi/kimi-k2.5

Hit Apply.

Step 4: add your Kimi API key

Go to platform.moonshot.ai and create an API key. Copy it.

Back in the admin dashboard, paste the key into the KIMI_API_KEY field. Hit Apply.

The dashboard should show nine models configured for Kimi.

Step 5: run Claude Code

fcc-claude

That's it. Claude Code launches with Kimi K2.5 as the model. You can verify by checking the model name in the status bar.

VS Code and JetBrains setup

If you use the Claude Code extension in VS Code, add these to your settings.json:

"claudeCode.environmentVariables": [
  { "name": "ANTHROPIC_BASE_URL", "value": "http://localhost:8082" },
  { "name": "ANTHROPIC_AUTH_TOKEN", "value": "freecc" },
  { "name": "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY", "value": "1" },
  { "name": "CLAUDE_CODE_AUTO_COMPACT_WINDOW", "value": "190000" }
]

For JetBrains, the same environment variables go into the ACP agents config file. Check the repo README for the exact path on your OS.

Switching providers

The admin dashboard lets you swap providers without touching code. If you want DeepSeek instead of Kimi, paste in a DeepSeek key and change the model string. Ollama for local inference, NVIDIA NIM for free cloud models, OpenRouter for the widest model selection.

You can also route different tiers to different providers. For example: Opus-tier requests to DeepSeek, Sonnet-tier to Kimi, Haiku-tier to a local Ollama model. Set MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU independently in the dashboard.

Honest limitations

This is not Claude. Kimi K2.5 is capable for everyday coding, scripts, and prototyping. For complex multi-file refactors and heavy agentic sessions, the gap between Kimi and Claude Opus is real. Free is free, not magic.
Tool use support varies. Some providers and models handle Claude Code's tool use better than others. Kimi is one of the better ones because of the native Anthropic protocol support.
Provider quotas still apply. The proxy is free. The upstream provider may have rate limits or usage caps on their free tier.
No stable releases. Install is from Git HEAD. Things can break between updates.
Context window is smaller. The proxy auto-sets CLAUDE_CODE_AUTO_COMPACT_WINDOW=190000 because non-Claude models typically have smaller context windows than Claude's 200k.

Who this is for

If you like the Claude Code interface but want to test other models without changing your workflow, this is a clean way to do it. If you're a student or hobbyist and can't justify $100/month for Claude Pro, Kimi K2.5 through this proxy covers a lot of ground.

If you need Claude-quality reasoning for production work, pay for Claude. The proxy is a free alternative, not an equivalent one.

Repo: github.com/Alishahryar1/free-claude-code