Skip to content
Inferam Docs

Messages

POST https://api.inferam.ai/v1/messages is Anthropic’s Messages format, which Claude Code and DeepSeek Harness use. Inferam translates it to the model server’s Chat Completions format and back, including streaming events.

Terminal window
curl https://api.inferam.ai/v1/messages \
  -H "x-api-key: $INFERAM_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4.1-Flash-1017",
    "max_tokens": 1024,
    "system": "You are a terse senior engineer.",
    "messages": [{"role": "user", "content": "What does a WAL protect against?"}]
  }'

Authorization: Bearer works here too. The anthropic-version and anthropic-beta headers are accepted and ignored.

Supported fields

Section titled “Supported fields”

model, messages, system (a string or text blocks), max_tokens, stream, stop_sequences, temperature, top_p, tools, tool_choice (auto, any, tool, none, and disable_parallel_tool_use), thinking and output_config.effort.

  • Content blocks: text, image (base64 source only), tool_use, tool_result and thinking. Other block types get a 400.
  • Tools: client tools with an input_schema. Anthropic server tools such as web search don’t run on Inferam.

Thinking

Section titled “Thinking”

The model thinks by default, and its reasoning comes back as thinking blocks before the answer. Their signature is always an empty string: Inferam has no signatures to give. Send thinking blocks back unchanged in assistant turns that made tool calls; the model expects its own reasoning on those turns.

Request Effect
"thinking": {"type": "disabled"} Thinking off
"thinking": {"type": "enabled", "budget_tokens": 4000} Reasoning capped at 4,000 tokens
"output_config": {"effort": "high"} Reasoning effort: low, medium, high, xhigh or max

Streaming

Section titled “Streaming”

With "stream": true you get the standard events: message_start, then content_block_start, content_block_delta (thinking_delta, text_delta, input_json_delta) and content_block_stop, then message_delta with the stop reason and usage, then message_stop. A failure after streaming has started is sent as an error event.

Stop reasons: end_turn, max_tokens and tool_use. A matched stop sequence ends the turn as end_turn, and stop_sequence is always null.

Usage

Section titled “Usage”
"usage": {
  "input_tokens": 256,
  "cache_read_input_tokens": 1792,
  "cache_creation_input_tokens": 0,
  "output_tokens": 312
}

As in Anthropic’s API, input_tokens excludes the tokens read from cache. The total input is input_tokens + cache_read_input_tokens.

Counting tokens

Section titled “Counting tokens”

POST /v1/messages/count_tokens takes the same body and returns {"input_tokens": N}. It’s a fast estimate (about four characters per token) that doesn’t call the model and isn’t metered. Use it for budgeting, not exact accounting.