API reference

Messages

POSThttps://api.offmem.app/v1/messages

Generates a model response in the Anthropic Messages format. Works with the official Anthropic SDKs and with Claude Code.

pip install anthropic

curl https://api.offmem.app/v1/messages \
  -H "x-api-key: $OFFMEM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "offmem-glm-5.3-abliterated",
    "max_tokens": 1024,
    "system": "Answer in English.",
    "messages": [{"role": "user", "content": "Explain what an API is in one sentence."}]
  }'

Headers

x-api-keyom_...required: API key. Authorization: Bearer also works.
Content-Typeapplication/jsonrequired: JSON body.
anthropic-versionstring: Accepted and not required.

Body

modelstringrequired: Model ID, or a claude-* name (see Accepted IDs). The response repeats the name you sent.
max_tokensintegerrequired: Output token limit, 1 or more.
messagesarrayrequired: Turns with role user or assistant. Messages with role system are merged into system.
systemstring | array: Instructions, as text or text blocks.
temperaturenumber: Forwarded to the model.
top_pnumber: Forwarded to the model.
stop_sequencesstring[]: Up to 4 sequences; extras are ignored.
toolsarray: Custom tools (name, description, input_schema). Anthropic server tools, such as web search, are ignored.
tool_choiceobject: auto, any, tool (with name) or none.
streamboolean: Returns Anthropic server-sent events.

Other fields, such as top_k, metadata and thinking, are ignored.

Content blocks

Block	Support
`text`	Supported.
`image`	Base64 (`source.type: base64`) or URL (`source.type: url`), only on models with vision.
`tool_use`	Supported, in assistant turns.
`tool_result`	Supported, with text content. `is_error: true` is sent to the model as an error.
`thinking`	Ignored.
`document`	Not supported: `400`. Send the text instead.

Response

200 OK

{
  "id": "msg_...",
  "type": "message",
  "role": "assistant",
  "model": "offmem-glm-5.3-abliterated",
  "content": [{ "type": "text", "text": "..." }],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 24,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "output_tokens": 40
  }
}

stop_reason	When
`end_turn`	The model finished the answer.
`max_tokens`	Reached `max_tokens`.
`tool_use`	The model called a tool.
`refusal`	The answer was blocked by a content filter.

usage.input_tokens excludes cached tokens, which come in cache_read_input_tokens. cache_creation_input_tokens is always 0. The request ID comes in the request-id header.

With "stream": true, the events follow the Anthropic order: message_start, content_block_start, content_block_delta (text_delta or input_json_delta), content_block_stop, message_delta and message_stop. A failure in the middle of the stream arrives as an error event.

Count tokens

POSThttps://api.offmem.app/v1/messages/count_tokens

Returns {"input_tokens": n}, an estimate of 3 characters per token over system, messages and tools. Free; requires a valid key.

cURL

curl https://api.offmem.app/v1/messages/count_tokens \
  -H "x-api-key: $OFFMEM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "offmem-glm-5.3-abliterated", "messages": [{"role": "user", "content": "ok"}]}'