/ api-reference / messages

Messages (Anthropic)

The Anthropic-compatible Messages endpoint — native protocol for Claude SDKs and Claude Code.

POST /v1/messages

The Anthropic Messages API, served by the base/mnt.ai router. Point any Anthropic SDK — or Claude Code itself — at https://api.basemnt.ai and model calls route through base/mnt.ai with the same auth, metering, and failover as the OpenAI-compatible surface.

Request

POST/v1/messages

curl https://api.basemnt.ai/v1/messages \
  -H "x-api-key: $SK_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.2",
    "max_tokens": 1024,
    "system": "You are a terse coding assistant.",
    "messages": [
      { "role": "user", "content": "Write a bash function that waits for a port." }
    ]
  }'

Body parameters

Parameter	Type	Default	Description
`model`	`string`	—	Required. Model ID from the catalog.
`messages`	`array`	—	Required. Turns of the conversation, alternating `user` / `assistant`.
`max_tokens`	`integer`	—	Required. Cap on generated tokens.
`system`	`string \| array`	—	System prompt.
`stream`	`boolean`	`false`	Stream events via SSE.
`temperature`	`number`	`1`	Sampling temperature, 0–1.
`tools`	`array`	—	Tool definitions (name, description, input schema).
`tool_choice`	`object`	—	`{"type":"auto"}`, `{"type":"any"}`, or `{"type":"tool","name":...}`.
`stop_sequences`	`array`	—	Sequences that end generation.

Response

200-response.json

{
  "id": "msg_01...",
  "type": "message",
  "role": "assistant",
  "model": "z-ai/glm-5.2",
  "content": [
    { "type": "text", "text": "wait_for_port() { ... }" }
  ],
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 48,
    "output_tokens": 172
  }
}

Content blocks mirror the Anthropic shape: text blocks, tool_use blocks with id, name, and input, and thinking blocks on reasoning-capable models.

Streaming

With stream: true, the endpoint emits the standard Anthropic event sequence — message_start, content_block_start, content_block_delta (with text_delta or input_json_delta), content_block_stop, message_delta, message_stop:

sse-event.json

{
  "type": "content_block_delta",
  "index": 0,
  "delta": { "type": "text_delta", "text": "wait_" }
}

Official Anthropic SDKs consume this natively — no code changes beyond the base URL.

Errors

Anthropic-style error bodies, same statuses as Chat Completions:

401-response.json

{
  "type": "error",
  "error": { "type": "authentication_error", "message": "Invalid API key" }
}

Why two protocols?

Tooling is sticky. Claude Code and Anthropic SDKs speak Messages; OpenAI SDKs and most agents speak Chat Completions. Rather than translate in your code, the router speaks both natively and routes to the same models either way — see connecting agents.