Chat

OpenAI Chat Completions compatible. Works with every model whose mode is chat. Supports streaming, tools, JSON output, images (vision models) and reasoning_effort (reasoning models). Unknown fields are passed to the upstream provider as-is, except the gateway-controlled fields listed below.

Billing. Before the call, the worst-case cost (prompt estimate + output limit) is reserved against your balance and the key's credit limit. If your credit can't cover the requested max_tokens / max_completion_tokens (or the default 4,096 when unset), the output limit is lowered to what you can afford; below 16 affordable tokens the request is refused with 402. After the call you are billed the actual usage and the rest of the reservation is returned.

Streaming. With "stream": true the response is text/event-stream with OpenAI chat.completion.chunk events ending in data: [DONE]. The gateway always sets stream_options.include_usage, so the last chunk carries usage.

Removed or refused by the gateway. Upstream-control fields are silently dropped: api_key, api_base, base_url, api_version, api_type, organization, user_config, model_list, mock_*, fallbacks, context_window_fallbacks, num_retries, max_retries, timeout, caching, cache, proxy_server_request, secret_fields, guardrails, extra_headers, headers, custom_llm_provider, deployment_id, base_model, model_info, *_cost_per_*, and anything prefixed litellm, aws_, vertex_, azure_ or watsonx_; metadata keys starting with litellm / model_ and metadata.tags are dropped too. Requests are refused (400) when they use provider-hosted tools (web_search*, web_fetch*, code_interpreter*, code_execution*, mcp*, image_generation*, computer_use*, file_search*, local_shell*) or web_search_options, reference objects stored upstream (file_id, vector_store_ids, item_reference, a string container, top-level conversation / prompt), contain more than 64 URLs, or contain URLs that resolve to private / internal addresses. Inline data: URLs are fine. If the key has redact secrets enabled, credentials found in the prompt are masked before forwarding.

OpenAI Responses API compatible. Works with every chat model: models whose provider has no Responses route are served transparently through Chat Completions (including streaming and tools), and the result is converted back to the Responses format.

previous_response_id is accepted only for response ids previously returned to your own account by this endpoint; any other id returns 404. Billing, output-limit capping, stripped fields and refused tools / references are the same as for Chat Completions (POST /v1/chat/completions); here the output limit field is max_output_tokens.

Streaming ("stream": true) returns the Responses event stream (response.created, …, response.completed).

Anthropic Messages API compatible, for Anthropic SDKs and Claude Code. Works with every chat model, not only Claude: the request is translated to Chat Completions and the result (including the stream) is translated back.

Supported: 'system (string or text blocks; cache_control is kept for claude* models),' text / image / document / tool_use / tool_result blocks, client tools with input_schema, tool_choice (auto, any, none, tool, disable_parallel_tool_use), max_tokens, temperature, top_p, stop_sequences, stream, and thinking (enabled with budget_tokens, or adaptive) on reasoning models — mapped to reasoning_effort (budget_tokens → low / medium / high). Other fields are ignored; Anthropic server tools (without input_schema) are dropped.

Dated model names such as claude-haiku-4-5-20251001 fall back to the undated catalog slug when only that one exists.

Errors use the Anthropic shape. Credit errors are 402 with type billing_error.

Claude Code:

Code
export ANTHROPIC_BASE_URL="https://ai.brandkamu.id"
export ANTHROPIC_AUTH_TOKEN="$API_KEY"
export ANTHROPIC_MODEL="claude-sonnet-5"
claude

Anthropic-compatible token counting. Returns a rough estimate (about 4 characters per token over system, messages and tools); it is not billed, but the call still requires a valid key and counts toward rate limits.