Skip to content

cursorpipe

Async Python client and OpenAI-compatible HTTP server for the Cursor Agent CLI — pipe prompts to frontier LLMs via ACP (Agent Client Protocol) with persistent sessions, streaming, and per-call model selection.

Three ways to use cursorpipe

Docker — self-hosted OpenAI-compatible API

Turn your Cursor subscription into a self-hosted LLM API with one command. Works with any language, any framework, any tool that speaks the OpenAI protocol.

git clone https://github.com/Abhi5h3k/cursorpipe.git && cd cursorpipe
export CURSOR_API_KEY=crsr_your_key_here
docker compose up

git clone https://github.com/Abhi5h3k/cursorpipe.git; cd cursorpipe
$env:CURSOR_API_KEY = "crsr_your_key_here"
docker compose up

HTTP Server — standalone

pip install "cursorpipe[server] @ git+https://github.com/Abhi5h3k/cursorpipe.git@v0.1.4"
cursorpipe-server

Python Library — async-first

from cursorpipe import CursorClient

client = CursorClient()
response = await client.generate(
    model="claude-4.5-sonnet-thinking",
    prompt="Explain what an API is in two sentences.",
)

Highlights

  • Language-agnostic HTTP API — OpenAI-compatible; works with any SDK, any language
  • Docker-ready — one docker compose up and you have an LLM API
  • Async-first Python library — built on asyncio for non-blocking LLM calls
  • Persistent ACP transport — keeps a single agent process alive, ~50ms overhead per request
  • Multi-turn sessions — server-side conversation history, no need to resend messages
  • Session isolation — every request gets a fresh session; no history leaks between users or calls
  • Warmup support — pre-start the process and pre-create sessions to eliminate cold-start latency
  • Per-call model selection — route different tasks to different models in a single client
  • Streaming — SSE over HTTP, or async for in Python
  • Auto-fallback — tries ACP first, falls back to subprocess if needed
  • Typed everything — Pydantic models, custom exceptions, py.typed for IDE support

Next steps

  • Getting Started — install and run your first prompt
  • HTTP Server — OpenAI-compatible API documentation
  • Docker — one-command Docker deployment
  • Examples — runnable scripts for every feature
  • API Reference — full method and class documentation
  • Architecture — how ACP, subprocess, and HTTP transports work
  • v1 vs v2 — feature comparison and which version to choose