Use Claude Code with the gateway

Claude Code talks to the LLM Gateway 2.0 /anthropic/v1/messages endpoint, and any alias can serve it, including GPT, Gemini, and Domino-hosted models. Grants, budgets, and guardrails apply as they do to other traffic. The usage dashboard can group each session by its claude-code-session tag.

How requests reach the model

Anthropic aliases, and Claude on Bedrock or Vertex, receive requests unchanged, so prompt caching markers, extended thinking, and beta headers work. Every other alias gets the request translated to OpenAI Chat Completions. Tool use, vision, and multi-turn conversations work. Prompt caching markers, extended thinking, and Anthropic beta headers are dropped.

Set up every Workspace in a Compute Environment

If you own the Compute Environment, add this to its Pre Run Script. Every Workspace built from it starts with Claude Code pointed at the gateway, and the gateway tags requests with the Project name.

cat > "$HOME/.anthropic-headers.sh" <<'EOF'
if [ -n "$DOMINO_PROJECT_NAME" ]; then
  export ANTHROPIC_CUSTOM_HEADERS="X-LLM-Tag-projectname: $DOMINO_PROJECT_NAME"
fi
export CLAUDE_CODE_ATTRIBUTION_HEADER=0
export ANTHROPIC_BASE_URL="https://<your-domino-host>/apps/<gateway-app-id>/anthropic"
EOF

CODE_DIR="${DOMINO_WORKING_DIR:-/mnt}"
mkdir -p "$CODE_DIR/.claude"
cat > "$CODE_DIR/.claude/settings.local.json" <<'EOF'
{
  "apiKeyHelper": "curl -sf http://localhost:8899/access-token"
}
EOF

MARKER="# >>> domino anthropic headers >>>"
if ! grep -qF "$MARKER" "$HOME/.bashrc" 2>/dev/null; then
  cat >> "$HOME/.bashrc" <<EOF

$MARKER
[ -f "\$HOME/.anthropic-headers.sh" ] && . "\$HOME/.anthropic-headers.sh"
# <<< domino anthropic headers <<<
EOF
fi

Start a Workspace, cd into the Project directory, and run claude --model <alias-name>. Launch from inside the Project directory, or Claude Code won’t find settings.local.json. The script rewrites .claude/settings.local.json on every Workspace start. Add it to .gitignore in Git-based Projects, and put personal Claude Code settings in .claude/settings.json instead.

Set up a single Workspace

Without access to the Compute Environment, write the same settings to your user config:

mkdir -p "$HOME/.claude"
cat > "$HOME/.claude/settings.json" <<EOF
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://<your-domino-host>/apps/<gateway-app-id>/anthropic",
    "ANTHROPIC_CUSTOM_HEADERS": "X-LLM-Tag-projectname: ${DOMINO_PROJECT_NAME}",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  },
  "apiKeyHelper": "curl -sf http://localhost:8899/access-token"
}
EOF

Use the gateway from outside Domino

From a laptop or a continuous integration (CI) system, the sidecar isn’t available. Use a Domino API key:

export ANTHROPIC_BASE_URL=https://<your-domino-host>/apps/<gateway-app-id>/anthropic
export ANTHROPIC_AUTH_TOKEN=<your-domino-api-key>
export CLAUDE_CODE_ATTRIBUTION_HEADER=0
claude --model <alias-name>

Before you roll it out

Register a Haiku alias. Claude Code makes background calls for titles and summaries to its default Haiku model by its Anthropic name. Without an alias of that exact name, those calls return 404 and titles and summaries degrade.
Keep CLAUDE_CODE_ATTRIBUTION_HEADER=0. Without it, the system prompt changes on every turn and prompt caching never hits.
Size alias ceilings for many small requests. An alias ceiling that runs out stops Claude Code mid-session. To limit extended-thinking spend, set LLM_GATEWAY_MAX_THINKING_BUDGET_TOKENS.
Tune timeouts and retries. Each alias retries on its own, and Claude Code retries up to 10 more times on top. For slow self-hosted models, raise the alias’s per-attempt timeout (default 60 seconds). To fail faster, lower its Max Retries.
Keep request bodies small. The Domino App proxy rejects request bodies larger than about 25 MB with a 413.
Pass tags explicitly from the Anthropic Python software development kit (SDK). The SDK doesn’t read ANTHROPIC_CUSTOM_HEADERS. Pass tags with default_headers on the client.

Control access, spend, and usage: set alias ceilings and grant budgets for Claude Code traffic, and break down spend by session or Project tag.

Core Concepts

Features

Use Claude Code with the gateway

How requests reach the model

Set up every Workspace in a Compute Environment

Set up a single Workspace

Use the gateway from outside Domino

Before you roll it out

​How requests reach the model

​Set up every Workspace in a Compute Environment

​Set up a single Workspace

​Use the gateway from outside Domino

​Before you roll it out

​Related

How requests reach the model

Set up every Workspace in a Compute Environment

Set up a single Workspace

Use the gateway from outside Domino

Before you roll it out

Related