anyllm 0.2.3

A thin, unified LLM abstraction layer. Call any LLM with a single API.

anyllm

Local-first LLM abstraction — one API for Ollama, llama.cpp, OpenAI, Anthropic, and HuggingFace.

anyllm is a lightweight abstraction layer over the most popular LLM providers. Unlike heavier alternatives, it is local-first: if Ollama is running on your machine, anyllm.chat("hello") just works — no API keys, no cloud. It also supports llama.cpp, OpenAI, Anthropic, and HuggingFace Transformers behind the same tiny API, with first-class support for tool/function calling, streaming, structured JSON outputs, multi-modal inputs, embeddings, and conversation memory.

Built by Viet-Anh Nguyen at NRL.ai.

Why anyllm?

• One-liner API — anyllm.chat("Hello") auto-detects your best local provider
• Plugin architecture — Add custom providers via @register_provider
• Local-first — Defaults to Ollama if available, no API key required
• Minimal core deps — Only httpx and pydantic; every provider is optional
• Production-ready — Streaming, async, tool-calling, retries, structured outputs

Installation

pip install anyllm

For optional providers:

pip install anyllm[openai]          # OpenAI GPT-4, GPT-3.5
pip install anyllm[anthropic]       # Claude 3.5 Sonnet / Opus / Haiku
pip install anyllm[llamacpp]        # llama.cpp local quantized models
pip install anyllm[transformers]    # HuggingFace Transformers (local)
pip install anyllm[all]             # everything

Ollama needs no Python package — just have it running at http://localhost:11434.

Python 3.8+ supported (tested on 3.8, 3.9, 3.10, 3.11, 3.12, 3.13)

Quick Start

import anyllm

# 1. Simple chat (auto-selects Ollama if running, else first configured provider)
reply = anyllm.chat("Explain RAG in one sentence.")
print(reply)

# 2. Specify a provider + model explicitly
reply = anyllm.chat(
    "What is the capital of France?",
    provider="ollama",
    model="llama3.1:8b",
)

# 3. Streaming (yields tokens as they are generated)
for chunk in anyllm.stream("Write a haiku about Python"):
    print(chunk, end="", flush=True)

# 4. Structured output (JSON mode — validates against a Pydantic model)
from pydantic import BaseModel
class Recipe(BaseModel):
    name: str
    ingredients: list[str]
    steps: list[str]

recipe = anyllm.chat("Give me a pasta recipe", response_model=Recipe)
print(recipe.name, recipe.ingredients)

Models & Methods

Providers (local-first priority)

Priority	Provider	How it works	Install
1	Ollama	HTTP client to http://localhost:11434 (default if reachable)	built-in
2	llama.cpp	Loads GGUF models via llama-cpp-python	anyllm[llamacpp]
3	OpenAI	REST API (gpt-4o, gpt-4o-mini, gpt-3.5-turbo)	anyllm[openai]
4	Anthropic	REST API (claude-3-5-sonnet, claude-3-5-haiku, claude-3-opus)	anyllm[anthropic]
5	HuggingFace Transformers	Loads any HF causal-LM model locally	anyllm[transformers]

Provider priority can be overridden via anyllm.set_priority([...]) or per-call with provider="...".

Features

• Tool / function calling — Pass Python functions; parameter schemas are auto-extracted from type hints and docstrings. Dispatches to Ollama tools, OpenAI tools, or Anthropic tool use automatically.
• Streaming — Unified token streaming for every provider (yields strings).
• Async — anyllm.achat(...), anyllm.astream(...).
• Structured outputs — response_model=MyPydanticModel uses native JSON mode on OpenAI/Anthropic/Ollama, falls back to regex extraction + retries elsewhere.
• Multi-modal — Pass images via anyllm.chat([..., {"image": "cat.jpg"}], model="gpt-4o").
• Embeddings — anyllm.embed("text", model="nomic-embed-text") with Ollama / OpenAI / sentence-transformers.
• Conversation memory — Conversation() with sliding-window history and optional disk persistence.
• Retries + timeouts — Configurable exponential backoff on transient errors.

API Reference

Function	Purpose
anyllm.chat(messages, **opts)	Chat completion -> str or Pydantic model
anyllm.stream(messages, **opts)	Generator yielding token chunks
anyllm.achat / astream	Async variants
anyllm.embed(text, model=...)	Returns list[float] embedding
anyllm.tools(fns, prompt)	Tool-calling loop with auto-dispatch
anyllm.Conversation(system=...)	Multi-turn memory
anyllm.list_models(provider=...)	Enumerate available models
anyllm.register_provider(name, cls)	Add a custom provider

CLI Usage

anyllm chat "Summarize this file" --file notes.txt
anyllm chat "Hi" --provider ollama --model llama3.1:8b
anyllm stream "Write a poem"
anyllm embed "hello world" --model nomic-embed-text
anyllm list-models --provider ollama

Examples

Tool calling with auto-extracted schemas

import anyllm

def get_weather(city: str, units: str = "celsius") -> dict:
    """Get the current weather for a city."""
    # ... call a weather API ...
    return {"city": city, "temp": 22, "units": units}

# anyllm inspects the signature + docstring, builds the JSON schema,
# runs the LLM, dispatches the tool call, and returns the final reply.
reply = anyllm.tools([get_weather], "What's the weather in Hanoi?")
print(reply)

Multi-turn conversation with memory

from anyllm import Conversation

conv = Conversation(system="You are a helpful Python tutor.", model="llama3.1:8b")
conv.send("What is a decorator?")
conv.send("Show me an example")          # remembers previous context
conv.save("chat.json")                   # persist to disk

Vision input with a multi-modal model

import anyllm

reply = anyllm.chat(
    [{"text": "What's in this image?"}, {"image": "cat.jpg"}],
    provider="openai",
    model="gpt-4o",
)

License

MIT (c) Viet-Anh Nguyen