Connect your coding agent

Claude Code

Point Claude Code at the gateway with ANTHROPIC_BASE_URL and authenticate with ANTHROPIC_AUTH_TOKEN, which is this gateway's sk- key. Leave ANTHROPIC_API_KEY unset, or the client may fall back to Anthropic's own authentication.

This gateway implements Anthropic's published gateway compatibility guide: anthropic-version and anthropic-beta are forwarded verbatim and treated as an open list, so a capability introduced by a newer client release is not stripped; upstream errors retain protocol structure with sensitive credentials filtered, and a cross-protocol route carries the capability_rejected: token that keeps the client's automatic recovery working; the system array and the cache_control markers inside it are forwarded unchanged, prompt caching still depends on model support; streams are relayed as they arrive and a keep-alive is emitted during upstream silence, reducing the risk of the client's 300-second watchdog timing out.

export ANTHROPIC_BASE_URL="https://www.link42.ai/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-..."
unset ANTHROPIC_API_KEY

# Copy an available text model id from GET /v1/models
export ANTHROPIC_MODEL="YOUR_TEXT_MODEL_ID"

set ANTHROPIC_BASE_URL=https://www.link42.ai/anthropic
set ANTHROPIC_AUTH_TOKEN=sk-...
set ANTHROPIC_API_KEY=
set ANTHROPIC_MODEL=YOUR_TEXT_MODEL_ID
claude

$env:ANTHROPIC_BASE_URL = "https://www.link42.ai/anthropic"
$env:ANTHROPIC_AUTH_TOKEN = "sk-..."
Remove-Item Env:ANTHROPIC_API_KEY -ErrorAction SilentlyContinue
$env:ANTHROPIC_MODEL = "YOUR_TEXT_MODEL_ID"
claude

Model discovery in Claude Code

With CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, Claude Code requests GET /v1/models and adds the result to its /model picker. It keeps only entries whose id contains claude or anthropic; that filter belongs to the client.

So Claude models your operator serves appear in the picker on their own, while the rest of this catalogue — DeepSeek, GLM, Seedance and the others — is callable over the Anthropic protocol but will not appear there, and is selected with ANTHROPIC_MODEL or /model instead. We do not rename a model to get it past that filter.

Codex

Codex speaks only the OpenAI protocol, and current configurations use the Responses API. Define a model_providers block in config.toml pointing at this gateway and select it with model_provider.

model_provider = "link42"
model = "YOUR_TEXT_MODEL_ID"

[model_providers.link42]
name = "LINK42"
base_url = "https://www.link42.ai/v1"
env_key = "LINK42_API_KEY"
wire_api = "responses"

Cursor and other OpenAI-compatible tools

Any OpenAI-compatible client that lets you set a base URL reaches the gateway at https://www.link42.ai/v1 with an sk- key. Both /v1/chat/completions and /v1/responses are served, so either wire protocol works.

Gemini CLI

Gemini CLI speaks the Gemini protocol: point GOOGLE_GEMINI_BASE_URL at the gateway's Gemini address, put this gateway's sk- key in GEMINI_API_KEY, and choose "Use Gemini API key" on first launch (signing in with a Google account bypasses the custom address). Pick the model with -m or GEMINI_MODEL, using an exact id from GET /v1/models.

Gemini CLI requires HTTPS for any address other than localhost; this gateway's address qualifies.

export GOOGLE_GEMINI_BASE_URL="https://www.link42.ai/gemini"
export GEMINI_API_KEY="sk-..."
export GEMINI_MODEL="YOUR_TEXT_MODEL_ID"
gemini

set GOOGLE_GEMINI_BASE_URL=https://www.link42.ai/gemini
set GEMINI_API_KEY=sk-...
set GEMINI_MODEL=YOUR_TEXT_MODEL_ID
gemini

$env:GOOGLE_GEMINI_BASE_URL = "https://www.link42.ai/gemini"
$env:GEMINI_API_KEY = "sk-..."
$env:GEMINI_MODEL = "YOUR_TEXT_MODEL_ID"
gemini

OpenCode

OpenCode connects through a custom provider: declare one that uses @ai-sdk/openai-compatible in opencode.json at the project root (or globally in ~/.config/opencode/opencode.json), set baseURL to https://www.link42.ai/v1, and read the key from the environment with {env:LINK42_API_KEY} instead of writing it into the file. List the model ids you use under models and pick one with /models.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "link42": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LINK42",
      "options": {
        "baseURL": "https://www.link42.ai/v1",
        "apiKey": "{env:LINK42_API_KEY}"
      },
      "models": {
        "YOUR_TEXT_MODEL_ID": { "name": "YOUR_TEXT_MODEL_ID" }
      }
    }
  }
}

export LINK42_API_KEY="sk-..."
opencode

$env:LINK42_API_KEY = "sk-..."
opencode

Cost per session and per subagent

Claude Code sends x-claude-code-session-id on every request, and requests from a subagent carry x-claude-code-agent-id and x-claude-code-parent-agent-id as well. The gateway records all three, so the console can total a session's spend and open it to show what each agent inside it cost — after a fan-out of subagents you can see which one spent the budget.

These identifiers come from the client and are used for attribution only, never as a credential or an identity. A request that arrives without them, such as an ordinary SDK call, is reported as unattributed rather than folded into a session.