InferKit - Deploy any AI function in 3 lines

Library for ML / Vision / LLM / Agent. No FastAPI boilerplate needed.

Install

pip install inferkit              # core (no Pillow)
pip install inferkit[vision]      # + Pillow for image in/out
pip install inferkit[torch,transformers]  # heavy ML stacks
pip install -e .[dev]             # local dev

Usage - 3 lines

# my_model.py
from inferkit import infer

@infer
async def run(payload, files=None):
    # payload: {"text": "..."}  files: list[bytes] for images/audio
    return {"output": f"echo: {payload.get('text')}"}

@infer.stream  # optional for LLM streaming
async def run_stream(payload):
    for tok in payload.get("text","").split():
        yield tok + " "

Run:

inferkit dev my_model.py --port 8000
# docs at http://localhost:8000/docs

Endpoints auto created:

• POST /api/v1/infer (multipart file + json)
• POST /api/v1/infer/json (json only)
• POST /api/v1/infer/stream (SSE)
• WS /ws/infer and WS /api/v1/ws/infer (WebSocket + streaming)

Image output helpers:

from inferkit import image_to_base64, bytes_to_response
from PIL import Image
return image_to_base64(Image.new("RGB",(512,512),"red"))
return bytes_to_response(png_bytes, "image/png")
# also still supported: return {"image_base64": b64} or return png_bytes

Init new project

inferkit init
# creates .env.example, .env, Dockerfile, my_model.py

Deploy (one command, any OS, auto detects Docker)

inferkit deploy
# if docker available -> docker compose/build
# else -> venv + uvicorn on INFERKIT_HOST:INFERKIT_PORT

Config via .env

INFERKIT_HOST=0.0.0.0
INFERKIT_PORT=8000
INFERKIT_CORS_ORIGINS=["*"]   # or * or http://a.com,http://b.com
INFERKIT_MAX_UPLOAD_MB=50
INFERKIT_RATE_LIMIT=60/minute
INFERKIT_API_KEY=            # if set, require X-API-Key header (also ?api_key=)
INFERKIT_DEBUG=false
# plain HOST/PORT/CORS_ORIGINS also work for backwards compat

Programmatic

from inferkit import serve
serve("my_model.py", port=8000)