SDK Limitations

cursorpipe v2 is built on the official cursor-sdk. The SDK does not expose all OpenAI API features. The table below documents the gaps.

Unsupported parameters are accepted in the request body and silently ignored so existing clients never receive unexpected 422 errors.

OpenAI parameters not supported

Parameter	SDK support	Notes
`temperature`	None	Accepted but ignored. The SDK does not expose sampling temperature.
`top_p`	None	Accepted but ignored.
`stop` sequences	None	Accepted but ignored.
`n` > 1 (multiple completions)	None	Only one completion per request. Would require N parallel agents.
`usage` (token counts)	None	SDK returns no token counts. `usage` field removed from responses.
`function_calling` / `tools`	None	The SDK handles tools internally. You cannot define custom tools.
`logprobs`	None	Not applicable to Cursor models.
`presence_penalty` / `frequency_penalty`	None	Accepted but ignored.
`logit_bias`	None	Accepted but ignored.
`seed`	None	Accepted but ignored.
`response_format`	None	Accepted but ignored. JSON mode is not guaranteed.

Endpoints not supported

Endpoint	Notes
`POST /v1/embeddings`	The Cursor SDK has no embedding API.
`POST /v1/images/generations`	Not available.
`POST /v1/audio/transcriptions`	Not available.
`POST /v1/fine-tuning/jobs`	Not applicable.

Other known differences

Feature	v2 behaviour
Token counts	Not available. The SDK does not return prompt or completion token counts. The `usage` field is omitted from responses.
`finish_reason`	Always `"stop"`. The SDK does not distinguish length cutoffs from normal completion.
`model` in response	The SDK resolves the model; if it returns a model ID, it is reflected in the response. Otherwise the requested model ID is echoed.
Concurrent requests	Each stateless request creates a temporary SDK agent. Stateful sessions share one agent per session. There is no global concurrency limit in cursorpipe — the SDK and Cursor API impose their own limits.
Rate limiting	When the Cursor API rate-limits a request, the server returns `429` with a standard OpenAI error body. The `Retry-After` header is not forwarded (SDK does not expose it).

Workarounds

Token counts needed? Estimate using tiktoken client-side:

import tiktoken

enc = tiktoken.encoding_for_model("gpt-4")
n_tokens = len(enc.encode(prompt))

Structured JSON output? Append an explicit instruction to your prompt:

"Respond with valid JSON only. No other text."

Multiple completions (n > 1)? Send N parallel requests:

import asyncio, httpx

async def complete(client, prompt):
    resp = await client.post("/v1/chat/completions", json={"model": "composer-2.5", "messages": [{"role":"user","content":prompt}]})
    return resp.json()["choices"][0]["message"]["content"]

async def n_completions(prompt, n=3):
    async with httpx.AsyncClient(base_url="http://localhost:8080") as c:
        return await asyncio.gather(*[complete(c, prompt) for _ in range(n)])