Claude Code

CLI dev agent from Anthropic. Connects over the native /v1/messages endpoint — your own key and catalog models.

Claude Code speaks the native Anthropic protocol: POST /v1/messages, and the gateway accepts the key both via the Authorization: Bearer header and via x-api-key. The ApiMira gateway answers this endpoint directly — set it up with five environment variables or a single settings file.

How to connect

1. 1Set ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN (the key is created in the dashboard and shown once), ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL and ANTHROPIC_SMALL_FAST_MODEL — in your terminal or in the Claude Code settings file.
2. 2So you do not have to set the variables again in every new terminal window, move them into the env block of the settings file ~/.claude/settings.json (on Windows that is C:\Users\Your name\.claude\settings.json) — the ready-made file is in the snippet below. The Claude Code extension for VS Code reads the same file: it launches the same CLI and never sees variables from a neighbouring terminal, so restart the editor after editing it. Settings for a single project go into .claude/settings.local.json in its root — that file stays out of git, so the key never reaches the repository.
3. 3Set ANTHROPIC_BASE_URL to the gateway address without /v1 — Claude Code appends /v1/messages itself, and a stray /v1 turns into a 404.
4. 4Set ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL and ANTHROPIC_SMALL_FAST_MODEL to our own Model IDs from the catalog. All three are required: the gateway does not recognize Anthropic's own model names and returns 404 on background tasks, even though the main conversation keeps working. ANTHROPIC_DEFAULT_HAIKU_MODEL is the current name for the background-model variable in Claude Code, ANTHROPIC_SMALL_FAST_MODEL is the old one (it usually still works); both are set so this does not depend on your Claude Code version.
5. 5Run claude — agentic mode works, including reading and editing files: the gateway supports tool calls. Streaming responses work too.

export ANTHROPIC_BASE_URL=https://apimira.com
export ANTHROPIC_AUTH_TOKEN=am-YOUR_KEY
export ANTHROPIC_MODEL=google/gemini-3.7-flash
export ANTHROPIC_DEFAULT_HAIKU_MODEL=google/gemini-3.7-flash
export ANTHROPIC_SMALL_FAST_MODEL=google/gemini-3.7-flash
claude

count_tokens returns an estimate, not an exact figure — use it as a guide rather than cutting your budget close to it; images are included in that estimate. Image input works: an image block with a base64 source (PNG, JPEG or WebP, up to 5 MB each, up to 50 per request) reaches the model, including an image returned by a tool — Claude Code reads an image file, and it arrives as a separate user turn. A URL instead of base64 (a source of type url) is not fetched, and a model without the "Sees images" mark answers with unsupported_capability. The endpoint rejects document blocks and the top_k parameter with a clear error, so attached files will not work in Claude Code. The thinking field and thinking blocks, on the contrary, are accepted and never passed on: there will be no extended thinking in the answer, but no error either — the MAX_THINKING_TOKENS variable is not needed. The whole prompt is capped at 400,000 characters (roughly 100k tokens), and some models add a context ceiling on top: crossing either limit returns an explicit error, not a silent cut; image bytes do not count towards the character cap. There is no cap of our own on max_tokens — Claude Code sets 32,000, and exactly that much is held on your balance for the answer, so keep a reserve. The cache_control field is accepted but ignored: manual cache markup is not supported. Caching still works automatically for models that list a cached-input rate on the Models page: when a prompt starts the same way as the previous one, those tokens are billed at the discounted rate, and their count arrives in usage.cache_read_input_tokens. This page's general rule that a parameter is never dropped silently is about the OpenAI-compatible /v1/chat/completions; here the request body is built from a whitelist of Anthropic fields, and anything outside it is exactly what gets dropped silently — cache_control above, for instance. Only unknown fields go quietly: document blocks and the top_k parameter come back as an error with code unsupported_parameter instead of disappearing from the request.

Function calling is supported, so agentic modes work. Strict JSON mode (response_format) and the retired functions format are rejected with an unsupported_parameter error — the parameter is never silently dropped. Image input is accepted by models marked "Sees images" on the Models page: a content part of type image_url with a data:image/png;base64,… URL (PNG, JPEG or WebP, up to 5 MB each, up to 50 per request). Remote URLs are not fetched, and a model without that mark answers with unsupported_capability — both errors name the reason outright.