Use Claude Code with CoralBricks

Point Claude Code (or any Anthropic SDK) at CoralBricks and run open models — GLM and Kimi with up to 1M context — with zero client changes.

The CoralBricks gateway speaks the Anthropic Messages API at POST /v1/messages. Anything built on Anthropic's SDKs — including Claude Code — can switch to CoralBricks inference with two environment variables.

Claude Code

export ANTHROPIC_BASE_URL="https://inference.coralbricks.ai"
export ANTHROPIC_AUTH_TOKEN="ak_..."     # your CoralBricks key
export ANTHROPIC_MODEL="glm-5.2-fp4"
export ANTHROPIC_SMALL_FAST_MODEL="glm-5.2-fp4"
claude

That's the whole setup. The agent loop — file edits, shell commands, tool calls, streaming — runs unmodified on the model you pick.

If you're logged into Claude Code with an Anthropic account, its stored credential can outrank ANTHROPIC_API_KEY — use ANTHROPIC_AUTH_TOKEN as shown (and unset ANTHROPIC_API_KEY) to make sure your CoralBricks key wins.

Anthropic SDKs

import anthropic

client = anthropic.Anthropic(
    base_url="https://inference.coralbricks.ai",
    api_key="ak_...",
)
message = client.messages.create(
    model="glm-5.2-fp4",
    max_tokens=1000,
    messages=[{"role": "user", "content": "Hello"}],
)

Both x-api-key (the SDK default) and Authorization: Bearer are accepted.

Available models

Model	Model ID	Context	Input $/M	Output $/M
GLM 5.2	`glm-5.2-fp4`	1M	$1.40	$4.40
Kimi K3	`kimi-k3`	1M	$3.00	$15.00
Kimi K2.6	`kimi-k2.6`	256K	$0.76	$4.00
GPT-OSS 120B	`gpt-oss-120b`	128K	$0.15	$0.60

Cached input tokens are always free — and they show up in the field Anthropic clients already read: usage.cache_read_input_tokens. Claude Code's own cost display reflects our cache economics directly.

What's supported

Non-streaming and streaming (full Anthropic SSE event stream, including fine-grained tool input_json_delta)
Tool use — client tools with input_schema, tool_choice (auto/any/tool/none), parallel tool calls
System prompts (top-level and mid-conversation system messages), stop sequences, temperature/top_p
POST /v1/messages/count_tokens (estimator-backed, deliberately biased slightly high)
anthropic-version/anthropic-beta headers and thinking/metadata/ cache_control params are accepted and ignored — prompt caching happens automatically in the serving tier

Not supported (clean errors, not silent drops)

Image and document content blocks (vision via this endpoint is coming)
Server tools (web search, code execution, computer use), MCP connector, containers — Claude Code's own tools all run client-side, so it is not affected
Extended thinking output — reasoning models think internally; reasoning is not returned as thinking blocks yet

The same key also works on the OpenAI-compatible surface — OpenCode, Cursor, Cline — and the full API reference is at /docs.