Claude Code

Point Claude Code at the gateway, multi-provider routing, and Anthropic release checklist.

Last updated: 2026-07-23

Claude Code talks Anthropic Messages (POST /v1/messages, SSE, tools, betas, count_tokens). Through the gateway you can:

Passthrough to Anthropic (API key or subscription OAuth), or
Translate to ChatGPT / SuperGrok / other OpenAI-compatible hosts while still using the Claude Code UI.

Start here for day-to-day use

Claude Code with ChatGPT, Claude & SuperGrok — full operator guide:

Subscription OAuth login (llm-gateway auth login …)
Any provider combo (gpt, grok, gpt+grok, multi, …)
Grok 4.5 + Composer 2.5 aliases
Shell helpers, settings.json, troubleshooting

This page is the Anthropic path checklist (release sign-off + minimal wiring). Prefer hermetic Go tests in CI; run the manual block for release candidates.

Minimal wiring (API key)

export ANTHROPIC_BASE_URL=http://localhost:8787
export ANTHROPIC_API_KEY=sk-ant-…   # or edge key when edge_auth is on
# optional: export ANTHROPIC_MODEL=claude-sonnet-4-20250514
claude
# or: ./examples/claude-code.sh

providers:
  anthropic:
    kind: anthropic
    base_url: "https://api.anthropic.com/v1"
    api_key_env: ANTHROPIC_API_KEY
defaults:
  anthropic_dialect: anthropic

With edge_auth enabled, set ANTHROPIC_API_KEY / KEY to an edge key and keep the upstream secret in api_key_env (or subscription oauth.credentials).

Automated coverage

go test ./proxy/ -count=1
go test ./ingress/anthropic/ ./egress/anthropic/ -count=1

Area	Tests (representative)
Messages passthrough	`TestAnthropicStreamPassthrough`, non-stream Anthropic server tests in `proxy/*`
Upstream errors	`TestAnthropicUpstreamError`, `TestAnthropicToOpenAIUpstreamErrorTranslated`
Translate path	`TestOpenAIToAnthropic`, `TestAnthropicToOpenAI`
count_tokens	`TestCountTokens*`, `proxy/count_tokens_test.go`
Edge auth open healthz	`TestEdgeAuthHealthzOpen`

Manual release sign-off

Config sketch

providers:
  anthropic:
    kind: anthropic
    base_url: "https://api.anthropic.com/v1"
    api_key_env: ANTHROPIC_API_KEY
  deepseek:
    kind: openai_compat
    base_url: "https://api.deepseek.com"
    api_key_env: DEEPSEEK_API_KEY
defaults:
  anthropic_dialect: anthropic
aliases:
  # cheap: deepseek/deepseek-chat

Functional checks

Non-stream short completion
Stream completion (tokens appear incrementally; no truncated SSE)
Tool use / function calling round-trip
Multimodal image input (if used in your workflows)
count_tokens still returns a number
Prompt-caching / beta headers: client anthropic-beta values reach upstream on passthrough
Model routing: bare id → anthropic default; provider/model and aliases
Translated path (optional): ANTHROPIC_MODEL=deepseek/deepseek-chat — full translation; stream input_tokens may be 0 until final event

Negative checks

Bad key → Anthropic-shaped error envelope
Unknown model provider → 404 dialect error
Gateway /healthz still 200 without credentials

Sign-off

go test ./... -race green
CHANGELOG entry if Anthropic surface or caveats changed
No secrets committed

Known limitations (translated path)

Topic	Behavior
Anthropic → openai_compat stream	`message_start.input_tokens` may be 0 until final usage
OpenAI-only fields	Dropped on translate — Deprecation policy
Files / Batch APIs	Not required for Claude Code chat
Media Contract Anthropic images	Planned; not required for text Claude Code