slimllm

Zero-dependency Python facade for OpenAI, Anthropic, and OpenRouter — designed for AWS Lambda.

Why

• No external dependencies — uses only Python stdlib (http.client, ssl, json, asyncio)
• LiteLLM-compatible API — same completion() call signature, swap without rewriting callers
• Lambda-friendly — tiny cold start, no bloat
• Supports streaming, tool use, and JSON mode

Supported providers

Install

pip install slimllm

Usage

import slimllm

# Non-streaming
resp = slimllm.completion(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.content)

# Streaming
stream = slimllm.completion("claude-3-5-sonnet-20241022", messages, stream=True)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="", flush=True)
final = stream.get_final_response()

# Async
resp = await slimllm.acompletion("gpt-4o", messages)

# Async streaming
async for chunk in slimllm.astream("claude-3-5-sonnet-20241022", messages):
    print(chunk.choices[0].delta.content or "", end="")

# Tool use (OpenAI format — auto-converted for Anthropic)
tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]
resp = slimllm.completion("claude-3-5-sonnet-20241022", messages, tools=tools)

# OpenRouter
resp = slimllm.completion("openrouter/meta-llama/llama-3.3-70b-instruct", messages)

API key resolution

Keys are looked up in this order:

1. Explicit api_key= kwarg
2. Environment variable (OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY)

Response shape

All providers return the same OpenAI-shaped types:

resp.content                          # str | None
resp.tool_calls                       # list[ToolCall] | None
resp.choices[0].finish_reason         # "stop" | "length" | "tool_calls"
resp.usage.prompt_tokens
resp.usage.completion_tokens

License

MIT