POST /v1/messages serves the Anthropic Messages API. A client written for the Anthropic SDKs or Claude Code works with PrivateMind once you change its base URL and key. It runs on the same models, keys, budgets and rate limits as chat completions, so use whichever format your code already speaks.
curl -s "https://api.privatemind.com/v1/messages" \
-H "x-api-key: $PMIND_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "Content-Type: application/json" \
-d '{
"model": "fast",
"max_tokens": 400,
"system": "You are a concise assistant.",
"messages": [
{"role": "user", "content": "Explain the CAP theorem."}
],
"temperature": 0.2
}'import anthropic
client = anthropic.Anthropic(
base_url="https://api.privatemind.com",
api_key="PMIND...:abcdef...",
)
message = client.messages.create(
model="fast",
max_tokens=400,
system="You are a concise assistant.",
messages=[{"role": "user", "content": "Explain the CAP theorem."}],
temperature=0.2,
)
print(next(b.text for b in message.content if b.type == "text"))import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic({
baseURL: 'https://api.privatemind.com',
apiKey: process.env.PMIND_KEY,
});
const message = await client.messages.create({
model: 'fast',
max_tokens: 400,
system: 'You are a concise assistant.',
messages: [{ role: 'user', content: 'Explain the CAP theorem.' }],
temperature: 0.2,
});
console.log(message.content.find((b) => b.type === 'text')?.text);The base URL has no /v1, because the Anthropic SDKs add it. The key goes in x-api-key or Authorization: Bearer (see Authentication). The SDKs send anthropic-version, and hand-written clients should too: it tells shared paths such as /v1/models and /v1/files to answer in this API's shape.
On a model server registered with No translation, the request goes to your vLLM as sent, apart from the model name, and its answer comes back unchanged. Parameters, thinking and usage then follow your vLLM's own Messages API support. The rest of this page describes the translated path.
Parameters
model(required): id or alias of a chat model fromGET /v1/models. Anthropic's ownclaude-*ids don't exist here, so use your catalog's model ids.max_tokens(required): maximum tokens to generate, thinking included.messages(required): alternatinguserandassistantturns.contentis a string or an array of blocks:text,image,document,tool_use,tool_result,thinking. Consecutive turns with the same role are merged. See Vision and Tool use.system: a string or an array oftextblocks, joined in order.temperature,top_p,top_k: sampling controls.stop_sequences: strings that end generation. A match reportsstop_reason: "stop_sequence"with the matchedstop_sequence.stream(defaultfalse): whentrue, the response arrives as Messages API events. See Streaming.tools,tool_choice: client tools with aninput_schema. A tool'sstrictflag is passed to the model server, and whether the call is held to the schema depends on the model's engine.tool_choicetakesauto,any,noneor{"type": "tool", "name": ...}, anddisable_parallel_tool_use. See Tool use.thinking,output_config.effort: how much the model thinks. See Thinking below.output_config.format:{"type": "json_schema", "schema": {...}}for structured output.metadata:user_idis passed to the model as the end-user id, and other keys are ignored. Unlikemetadataon chat completions, it isn't stored as usage tags.
A final assistant turn is a prefill: the model continues from its text. Fields the API does not use, such as cache_control, context_management and beta flags, are accepted and ignored, so SDK upgrades do not break calls.
Thinking
Reasoning models answer with thinking blocks before their text. You set how deep they think with thinking or output_config.effort, which map onto the same levels as reasoning_effort on chat completions:
| You send | Effort |
|---|---|
Nothing, or thinking: {type: "adaptive"} |
The model's default |
thinking: {type: "enabled", budget_tokens: N} |
low below 4096, medium below 16384, high from 16384 |
output_config: {effort: "low"}, "medium", "high" |
That level |
output_config: {effort: "xhigh"} or "max" |
max |
thinking: {type: "disabled"} |
off |
Runtime modes and levels work as for reasoning_effort, including a level missing from the model's capabilities.reasoning_effort_levels. A thinking-only model, such as Kimi K3.1 Pro or GLM 5.3, can't stop thinking, so it returns 400 for thinking: {type: "disabled"}.
On a model server you launched, the level reaches your vLLM as on chat completions.
thinking: {display: "omitted"} keeps the thinking blocks but empties their text.
Every thinking block carries a signature. Send thinking blocks back unchanged in later turns, as the SDKs do, and the earlier reasoning is passed back to the model. Whether the model uses it depends on its chat template: many keep only the reasoning of the current tool loop.
Response
{
"id": "msg_abc123",
"type": "message",
"role": "assistant",
"model": "fast",
"content": [
{ "type": "thinking", "thinking": "The user wants a short ...", "signature": "privatemind" },
{ "type": "text", "text": "The CAP theorem says ..." }
],
"stop_reason": "end_turn",
"stop_sequence": null,
"usage": {
"input_tokens": 42,
"output_tokens": 128,
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 0
}
}model echoes what you sent. Content comes thinking first, then text, then any tool_use blocks.
Usage and prompt caching
Prompt caching works as on chat completions: it is automatic and needs no cache_control. The Messages API reports it the Anthropic way. input_tokens counts only the part of the prompt that was not read from the cache, and cache_read_input_tokens counts the part that was. Their sum is the full prompt, the same number chat completions reports as prompt_tokens. cache_creation_input_tokens is always 0: there is no cache-write charge.
Stop reasons
| Value | Meaning |
|---|---|
end_turn |
Natural end of the answer |
max_tokens |
Hit max_tokens or the context window. If the cut falls inside a tool call, its tool_use input is incomplete or empty: don't run that call. Raise max_tokens and retry |
stop_sequence |
One of stop_sequences matched; see stop_sequence |
tool_use |
The model called one or more tools. See Tool use |
refusal |
Output was blocked by a content filter. The models this endpoint serves do not produce it today |
What is not supported
| Feature | Behaviour |
|---|---|
Server tools (web_search, web_fetch, code_execution, the MCP connector) |
Not run; the model answers without them |
| PDF documents | 400, so a PDF is never silently dropped. Text documents work |
| Vendor provider models (Claude served through a provider card) | 400 naming the dialect; use chat completions for those |
Related endpoints
| Endpoint | Page |
|---|---|
POST /v1/messages/count_tokens |
Count tokens |
GET /v1/models, GET /v1/models/{id} |
Models |
/v1/files |
Files |
/v1/messages/batches |
Message Batches |
Where next
- Anthropic SDK for client setup in each language.
- Streaming for the event stream.
- Coding agents to point Claude Code at PrivateMind.
- Errors for the error shape and status codes.