Claude Code

opencodex serves POST /v1/messages (plus count_tokens) alongside /v1/responses, so Claude Code can use every routed provider — OAuth logins, account pools, key failover and sidecars included — with zero extra auth work.

Quickstart

ocx claude

ocx claude ensures the proxy is running, then launches Claude Code with the environment wired:

Variable	Value
ANTHROPIC_BASE_URL	http://127.0.0.1:<port>
ANTHROPIC_AUTH_TOKEN	Only when the proxy requires an API key — otherwise it is NOT set, so your claude.ai login (subscription + connectors) stays active
CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY	1 (native /model picker discovery)
CLAUDE_CODE_AUTO_COMPACT_WINDOW	Auto-context compaction threshold (default 350000); only injected when auto-context is enabled
ANTHROPIC_MODEL	claudeCode.model (optional)
ANTHROPIC_DEFAULT_HAIKU_MODEL	claudeCode.tierModels.haiku ?? claudeCode.smallFastModel (optional; legacy ANTHROPIC_SMALL_FAST_MODEL too)
ANTHROPIC_DEFAULT_{OPUS,SONNET,FABLE}_MODEL	claudeCode.tierModels.* (optional)
CLAUDE_CODE_ALWAYS_ENABLE_EFFORT	1 when alwaysEnableEffort is on (conditional)
CLAUDE_CODE_MAX_CONTEXT_TOKENS / DISABLE_COMPACT	Legacy context override when maxContextTokens is set (conditional)
Variables you export yourself always win. Extra arguments pass through: ocx claude -p "hello".

System environment integration (macOS)

When claudeCode.systemEnv is set to true (default: off), ocx start uses launchctl setenv to inject ANTHROPIC_BASE_URL and the related Claude Code environment variables system-wide. New terminal windows and tabs therefore route plain claude commands through the proxy without requiring the ocx claude wrapper. Already-open shells are unaffected and must be reopened.

ocx stop and proxy shutdown unset the injected keys (it does not restore previous values — only the keys opencodex injected are removed). The proxy also writes ~/.opencodex/claude-env.sh; ocx start installs a .zshrc source hook that loads it automatically.

Disable with claudeCode.systemEnv: false in the configuration or with the GUI toggle. This feature is macOS-only; on other platforms, use ocx claude.

Native Claude passthrough (subscription pierce)

With no auth override set, Claude Code keeps its claude.ai OAuth login and sends it to the proxy. Requests for genuine claude*/anthropic* models that no alias or model map claims are forwarded verbatim to api.anthropic.com with your credential — betas, thinking signatures, prompt caching and billing identity stay fully native, and routed models keep working in the same session via the picker aliases.

Header handling: hop-by-hop headers plus host, content-length, accept-encoding, x-opencodex-api-key, and origin are stripped before forwarding. All other headers (including anthropic-beta and anthropic-version) pass through.

The passthrough fires when all four conditions are met: nativePassthrough is not false; the model begins with claude or anthropic; the bearer or x-api-key starts with sk-ant-; and alias/model-map resolution returns the same model unchanged. This also means the “claude.ai connectors are disabled” warning no longer appears with ocx claude.

Disable with claudeCode.nativePassthrough: false; point elsewhere with claudeCode.anthropicBaseUrl.

The /model picker (“From gateway”)

Claude Code 2.1.129+ discovers gateway models via GET /v1/models?limit=1000 and lists them in the native /model picker labeled “From gateway”. Because the picker only accepts ids beginning with claude or anthropic, opencodex exposes routed models as stable, reversible aliases:

Surface	Format	Example
Claude Code CLI	claude-ocx-<provider>--<model>	claude-ocx-native--gpt-5.6-sol
Claude Desktop 3P	claude-opus-4-8-<code> (3-char base36 hash)	claude-opus-4-8-ncb

The proxy picks the family per request: ?ids=cli or ?ids=desktop wins; otherwise the claude-code/* user-agent gets the readable CLI form and other clients get the Desktop hash. Both families decode forever — a model saved in settings.json under either form keeps working.

Alias grammar rules: provider must not contain / or -- or equal native; model must not contain /. Routes the readable form cannot express fall back to the hashed alias. Model ids MAY contain -- (resolution splits on the first -- only); native slugs containing -- fall back to the hashed form.

Model resolution order: [1m] marker stripped → readable alias decoded → Desktop hashed alias decoded → modelMap exact match → date-stripped match (-20250514 removed) → passthrough.

Each entry carries a display name like gemini-3-pro (gemini), plus full model capabilities (reasoning-effort ladder, thinking types) in the official ModelInfo shape. Real Anthropic models keep their canonical ids on both surfaces.

Context-variant [1m] marker

Models with an authoritative context window of 1M (or, under auto-context, above 200k and at least the compaction threshold) get an extra …[1m] picker row. Selecting it makes Claude Code account a full 1M context. The proxy strips the case-insensitive [1m] suffix before alias resolution and routing.

Auto context (big-context models without the 200k ceiling)

Claude Code accounts 200k tokens for any model it does not recognize. Auto context (on by default) fixes that:

1. Models whose real window is above 200k and at least the auto-compact threshold get the [1m] marker on their picker rows and env slots.
2. CLAUDE_CODE_AUTO_COMPACT_WINDOW (default 350000, range 100000–1000000) is injected so the conversation auto-summarizes at that point.

Three config states:

• absent / true: enabled (default)
• false: disabled — no markers, no compaction window injection
• legacy maxContextTokens set: auto-context is implicitly disabled

The compaction value is adjustable on the Claude page. Warning: raising it past a model’s real window breaks that model — the chat errors out before the summary can fire.

Sub-1M native Anthropic models are never auto-marked. Values you export yourself always win (the proxy uses YOUR value to decide which models are safe to mark). Invalid hand-edited config values fall back to 350k.

Effective model environment

effectiveModelEnv computes six slots injected by ocx claude / system env / shell file: ANTHROPIC_MODEL, four ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU,FABLE}_MODEL, and legacy ANTHROPIC_SMALL_FAST_MODEL. The effective Haiku is tierModels.haiku ?? smallFastModel, fed to both Haiku variables.

Roster agents (injectAgents)

ocx claude (and the system-env daemon) syncs your featured subagent roster (Subagents tab, up to 5 models) plus ocx-self into ~/.claude/agents/ocx-*.md.

• ocx-self pins your /model picker default (falling back to claudeCode.model); omitted when neither exists. It does NOT use model inheritance.
• Each agent body contains an <!-- ocx-route: <model> --> directive — the proxy uses this to pin the real route. The Agent tool’s model argument is therefore inert; pass "sonnet" as a placeholder.
• Frontmatter carries the alias; routing is directive-driven.
• Only marker-verified ocx-*.md files containing generated-by: opencodex are ever overwritten or pruned; your own agents are never touched.
• Files are atomically synced per file (write + rename).
• enabled: false or injectAgents: false prunes all verified-owned definitions.
• GUI PUT and roster changes resync immediately; launcher/system-env sync at launch.

Dispatch: subagent_type: "ocx-gpt-5-6-sol". 1M-capable targets carry [1m] automatically.

Bundled-skill elision (blockedSkills)

Claude Code’s bundled claude-api skill injects ~840KB (~136k tokens) of Anthropic documentation that auto-triggers on Claude model mentions. Routed models are not trained on that bundle, so by default opencodex replaces the skill’s content with a short stub on routed requests. Native Anthropic passthrough is untouched.

Two carriers are handled:

1. Tool-result carrier: assistant Skill(...) calls — the paired tool_result body is replaced by a stub when the lowercased JSON input contains a blocked name.
2. Text-block carrier: a user text block ≥10,000 characters starting with Base directory for this skill: — matched when the directory basename equals a blocked name (case-insensitive).

Configure with claudeCode.blockedSkills (default ["claude-api"]; [] disables elision entirely). The stub keeps tool call/result pairing intact.

Model map (interception)

claudeCode.modelMap rewrites inbound Anthropic model ids before routing:

{
  "claudeCode": {
    "modelMap": {
      "claude-sonnet-4-5": "gemini/gemini-3-pro",
      "claude-haiku-4-5": "gemini/gemini-3-flash"
    }
  }
}

Lookup order: discovery alias → exact id → id with date suffix stripped (-20250514) → passthrough.

Sidecar matrix: web search and image understanding

Routed models do not all have the same hosted tools or image support. opencodex fills those gaps before the main model answers:

• The web-search sidecar runs the real hosted search, then gives the routed model the answer and sources as a tool result.
• The vision sidecar describes an attached image before calling a model listed in noVisionModels, then replaces the image with that description.

Both sidecars can use either backend:

Backend	How it runs	What it requires
openai	A small GPT model through the ChatGPT forward provider	A ChatGPT login and an enabled authMode: "forward" provider
anthropic	Claude through stored Anthropic OAuth; web search uses web_search_20250305 and vision sends the image to Claude for description	An enabled adapter: "anthropic", authMode: "oauth" provider whose active stored account is not marked needsReauth

An explicit backend always wins. When it is omitted, opencodex selects anthropic if a usable stored Anthropic OAuth account exists; otherwise it selects openai. Explicitly selecting anthropic without a usable credential fails closed: opencodex does not silently borrow ChatGPT credentials or switch backends. The OpenAI backend likewise stays off without both login auth and a forward provider.

Claude-inbound routed replays attach the main ChatGPT login to the internal request, so OpenAI sidecars remain reachable even though Claude Code’s inbound bearer is only the proxy credential. That bearer is never forwarded to the routed main provider.

{
  "webSearchSidecar": {
    "backend": "anthropic",
    "model": "claude-sonnet-5",
    "maxSearchesPerTurn": 3
  },
  "visionSidecar": {
    "backend": "anthropic",
    "model": "claude-sonnet-5",
    "maxDescriptionsPerTurn": 8
  }
}

maxDescriptionsPerTurn limits new image descriptions in one main-model turn. Cache hits and duplicate in-flight descriptions do not consume the cap. Successful descriptions for data: images are cached by backend, model, detail, image bytes, and request context, so the same image-and-context pair is not described again on every replay. Remote https: images are never cached because their contents can change.

See the configuration reference for every key. Anthropic-OAuth web search and image description reuse the repository’s existing Claude Code OAuth fingerprint precedent, but should still be soak-tested with your account and workload before you depend on them for long unattended runs.

Reasoning effort

Claude Code’s /effort setting is preserved across the adapter:

Wire format	Mapping
thinking.type: "adaptive" + output_config.effort	Effort passed directly (minimal|low|medium|high|xhigh|max|ultra)
thinking.type: "enabled" + budget_tokens	≤4096→low, ≤16384→medium, above→high
thinking.type: "disabled"	Reasoning parameters omitted entirely

The resolved value appears in the request log’s Reasoning effort column.

Inbound translation (Messages → Responses)

The proxy translates every Anthropic Messages API request into the Codex Responses API format:

Messages input	Responses output
Top-level system	instructions (text blocks joined with \n\n)
messages[].role: "system"	Also folded into instructions
User text / image	input_text / input_image (base64 → data URL)
Assistant text	output_text
Assistant tool_use	function_call (input → JSON-stringified arguments)
User tool_result	function_call_output (is_error → [tool error] prefix)
thinking / redacted_thinking replay	Dropped
Function tools	{type: "function"} (web_search* → {type: "web_search"})
tool_choice	auto→auto, none→none, any→required, named→{type:"function",name}
max_tokens	max_output_tokens
stop_sequences	stop

Error cases (400): malformed JSON; missing/empty model; missing/empty messages; unsupported role; tool_result without tool_use_id; tool_use without id/name; named tool_choice without name.

Outbound translation (Responses → Messages SSE)

Responses event	Messages SSE
response.created	message_start + ping
Heartbeat	ping
Text deltas	content_block_start → content_block_delta (text) → content_block_stop
Reasoning summary/text	thinking block with synthetic signature
Function-call frames	tool_use block with input_json_delta
Terminal event	message_delta → message_stop
EOF before terminal	502-style api_error

Stop reason mapping: completed → tool_use (if any tool call) or end_turn; incomplete/max_output_tokens → max_tokens; incomplete/content_filter → refusal.

Error taxonomy: 400 invalid_request_error, 401 authentication_error, 402 billing_error, 403 permission_error, 404 not_found_error, 409 conflict_error, 413 request_too_large, 429 rate_limit_error, 504 timeout_error, 529 overloaded_error, other 5xx api_error. Retry-After is preserved.

Prompt caching and token usage

Anthropic-routed requests: the adapter manages cache breakpoints for tools, system content, and the penultimate user message, plus top-level automatic cache_control. Stable turns normally produce about a 99.9% cache hit rate.

Native OpenAI/ChatGPT routing: derives a session-scoped prompt_cache_key (from metadata.user_id when present, falling back to a system-content hash) and session_id header for cache affinity. The cache key includes model and full tool schemas.

Token math: Anthropic output subtracts cached_tokens and cache_write_tokens from input_tokens, exposing them as cache_read_input_tokens and cache_creation_input_tokens. Request logs map those back to inclusive inputTokens, with reads in both cachedInputTokens and cacheReadInputTokens, writes in cacheCreationInputTokens. The Usage page reports cache hits and cache creation separately.

count_tokens: routed models use an approximation (serialized system + messages + tools). Native Anthropic models with an sk-ant- credential pass the request through to the real Anthropic /v1/messages/count_tokens endpoint.

Debug capture

ocx debug claude on|off|status|reset, OCX_CLAUDE_DEBUG=1, or PUT /api/debug {"claude": true} controls inbound capture. GET /api/claude/inbound-debug returns {enabled, entries} (newest first, ring of 20).

Each entry records: at, endpoint, model, resolvedModel, stream, maxTokens, thinkingType, thinkingBudgetTokens, outputConfigEffort, metadataKeys, hasMetadataUserId, hasSystem, raw anthropicBeta, and eight-character HMAC equality tags for user id / system. No prompt text, raw object, or stable cross-run hash is stored. Disabling Claude debug immediately clears the ring.

GUI (Claude page)

The dashboard sidebar has a dedicated Claude page (below API) and a Claude ON toggle (label intentionally identical in every language). The page shows:

• Inbound kill switch (enabled toggle)
• Quickstart (ocx claude) and manual env block
• Fast Mode selector (Auto / ON / OFF)
• Auto-context toggle and compaction threshold dropdown
• Subagent auto-registration toggle
• Model interception (modelMap) editor
• Live preview of picker aliases

GET /api/claude-code returns effective defaults, config, context-window registry, effective env, available route ids, aliases, and port. PUT /api/claude-code is partial and preserves omitted fields; null resets context/blocklist/compact-window values.

Troubleshooting

Claude Code says “Did 0 searches” — Current builds translate completed Responses web_search_call items into paired Anthropic server_tool_use and web_search_tool_result blocks, including usage.server_tool_use.web_search_requests. Update opencodex if an older build completed the search but Claude Code still counted zero.

A sidecar does not activate — For backend: "openai", confirm you are logged into ChatGPT and have an enabled authMode: "forward" provider. For backend: "anthropic", confirm the active stored Anthropic OAuth account is not marked needsReauth. An explicit Anthropic selection without that credential intentionally fails closed.

“claude.ai connectors are disabled” — An ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN is set in your shell. ocx claude deliberately does NOT set ANTHROPIC_API_KEY; if you have it exported, unset it. ocx claude injects ANTHROPIC_BASE_URL, discovery, auto-context, and configured model slots — but never ANTHROPIC_API_KEY.

Models not showing in /model picker — Verify CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 is set (automatic with ocx claude). Run ocx claude to refresh the gateway model cache at ~/.claude/cache/gateway-models.json. Check claudeCode.enabled is not false.

Stale environment after port change — If the proxy port changed, old shells may have a stale ANTHROPIC_BASE_URL. Open a new terminal, or re-run ocx claude.

200k context ceiling despite big model — Select the [1m] variant in the picker, or enable auto-context (on by default). If the picker shows no [1m] row, the model’s authoritative context window may be below the auto-compact threshold.

High token count from skill loads — The bundled claude-api skill (~136k tokens) auto-loads on Claude model mentions. This is normal for native passthrough; on routed models, opencodex stubs it by default (blockedSkills: ["claude-api"]).

Subagent dispatches to wrong model — Roster agents (ocx-*) use <!-- ocx-route: ... --> directives, not the Agent tool’s model argument. Make sure the directive matches the intended route. Pass "sonnet" as the model placeholder.