Use KiteRouter as the gateway for coding agents by configuring each tool with the API shape it already speaks. Codex uses OpenAI Responses, Claude Code uses Anthropic Messages, and Gemini CLI uses the Gemini API.

Share this section with teammates: https://kiterouter.com/docs/coding-tools

Before you start

1. Create a KiteRouter API key from Dashboard -> API Keys.
2. Copy the exact model id from the Models page.
3. Restart the coding tool after changing env vars or config files.
4. Test with a tiny prompt before launching a long agent task.

Config file locations

Use the user-level file for API keys and personal provider settings. Project-level files can be committed for team defaults, but do not put kr- keys in a repository.

Codex

Codex reads config.toml from the Codex user directory above. Current Codex custom providers use the Responses API, so point base_url at KiteRouter's OpenAI-compatible versioned base URL and keep wire_api = "responses".

tomlCopy
1model = "openai/gpt-5.5"
2model_provider = "kiterouter"
3
4[model_providers.kiterouter]
5name = "KiteRouter"
6base_url = "https://kiterouter.com/api/v1"
7env_key = "KITEROUTER_API_KEY"
8wire_api = "responses"

macOS / Linux / WSL:

bashCopy
1export KITEROUTER_API_KEY="kr-your-api-key"
2codex exec "Reply with: ready"

Windows PowerShell:

powershellCopy
1$env:KITEROUTER_API_KEY = "kr-your-api-key"
2codex exec "Reply with: ready"

Use the model value for the KiteRouter model you want Codex to run. If you switch models often, set the provider once and override the model from the Codex CLI when needed.

Claude Code

Claude Code supports LLM gateways through ANTHROPIC_BASE_URL. Use ANTHROPIC_AUTH_TOKEN for KiteRouter so Claude Code sends your kr- key as a Bearer token.

macOS / Linux / WSL:

bashCopy
1export ANTHROPIC_BASE_URL="https://kiterouter.com/api"
2export ANTHROPIC_AUTH_TOKEN="kr-your-api-key"
3export ANTHROPIC_MODEL="anthropic/claude-sonnet-4-5"
4claude -p "Reply with: ready"

Windows PowerShell:

powershellCopy
1$env:ANTHROPIC_BASE_URL = "https://kiterouter.com/api"
2$env:ANTHROPIC_AUTH_TOKEN = "kr-your-api-key"
3$env:ANTHROPIC_MODEL = "anthropic/claude-sonnet-4-5"
4claude -p "Reply with: ready"

For a persistent setup, put the same values in the user settings.json file from the platform table:

jsonCopy
1{
2  "env": {
3    "ANTHROPIC_BASE_URL": "https://kiterouter.com/api",
4    "ANTHROPIC_AUTH_TOKEN": "kr-your-api-key",
5    "ANTHROPIC_MODEL": "anthropic/claude-sonnet-4-5"
6  }
7}

If you want Claude Code's /model picker to discover gateway models, enable gateway model discovery:

bashCopy
1export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

Gemini CLI

Gemini CLI supports a custom Gemini API base URL with GOOGLE_GEMINI_BASE_URL. Set the API version to v1 so requests land on KiteRouter's /api/v1/models/... routes.

macOS / Linux / WSL:

bashCopy
1export GEMINI_API_KEY="kr-your-api-key"
2export GOOGLE_GEMINI_BASE_URL="https://kiterouter.com/api"
3export GOOGLE_GENAI_API_VERSION="v1"
4export GEMINI_MODEL="gemini-2.5-pro"
5gemini -p "Reply with: ready"

Windows PowerShell:

powershellCopy
1$env:GEMINI_API_KEY = "kr-your-api-key"
2$env:GOOGLE_GEMINI_BASE_URL = "https://kiterouter.com/api"
3$env:GOOGLE_GENAI_API_VERSION = "v1"
4$env:GEMINI_MODEL = "gemini-2.5-pro"
5gemini -p "Reply with: ready"

For a persistent setup, put the env values in the Gemini .env file from the platform table:

dotenvCopy
1GEMINI_API_KEY=kr-your-api-key
2GOOGLE_GEMINI_BASE_URL=https://kiterouter.com/api
3GOOGLE_GENAI_API_VERSION=v1
4GEMINI_MODEL=gemini-2.5-pro

Then make sure Gemini CLI is set to API-key auth in the user settings.json file:

jsonCopy
1{
2  "security": {
3    "auth": {
4      "selectedType": "gemini-api-key"
5    }
6  },
7  "model": {
8    "name": "gemini-2.5-pro"
9  }
10}

Use a Gemini-family model id from the Models page.

CC Switch

If you manage several coding agents on the same machine, create one KiteRouter provider per protocol:

Keep the provider name tied to the protocol, not the model. That makes it obvious when a tool is about to send OpenAI, Anthropic, or Gemini-shaped requests.

Troubleshooting