anyllm 0.2.2

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

anyllm

Chat with any LLM — local-first, cloud-optional

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

Local-first: Uses Ollama/llama.cpp by default. Cloud APIs optional. When local models are available, anyllm automatically prefers them over cloud providers, ensuring your applications work offline.

Simpler than litellm -- focused on the essentials. No bloat, no complex abstractions. Just anyllm.chat().

Provider Support

Provider	API Key Env Var	Streaming	Local
OpenAI (+ compatible APIs)	OPENAI_API_KEY	Yes	No
Anthropic Claude	ANTHROPIC_API_KEY	Yes	No
Ollama	--	Yes	Yes
llama.cpp server	--	Yes	Yes

Installation

pip install anyllm

With optional provider SDKs:

pip install anyllm[openai]      # OpenAI SDK
pip install anyllm[anthropic]   # Anthropic SDK
pip install anyllm[all]         # All optional SDKs

Note: anyllm works without any provider SDKs installed -- it uses httpx for all HTTP calls by default.

Quick Start

Simple Chat

import anyllm

# Auto-detects available provider: checks Ollama (localhost:11434) first,
# then env vars (OPENAI_API_KEY, ANTHROPIC_API_KEY)
# Sends HTTP request via httpx — no SDK dependencies required
# Returns Response(content, model, usage) with token counts
response = anyllm.chat("What is the meaning of life?")
print(response.content)

OpenAI

export OPENAI_API_KEY="sk-..."

response = anyllm.chat("Hello!", model="openai/gpt-4")
print(response.content)
print(f"Tokens used: {response.usage.total_tokens}")

Anthropic Claude

export ANTHROPIC_API_KEY="sk-ant-..."

response = anyllm.chat("Hello!", model="anthropic/claude-sonnet-4-20250514")
print(response.content)

Ollama (Local)

ollama pull llama3

response = anyllm.chat("Hello!", model="ollama/llama3")
print(response.content)

llama.cpp Server

./llama-server -m model.gguf

response = anyllm.chat("Hello!", model="llamacpp/default")
print(response.content)

Streaming

for chunk in anyllm.chat("Tell me a story", model="openai/gpt-4", stream=True):
    print(chunk, end="", flush=True)
print()

Tool / Function Calling

# Tool/function calling — LLM decides when to call your functions
# anyllm forwards tool schemas to the provider, parses tool_call responses,
# and surfaces them in response.tool_calls for you to execute
response = anyllm.chat("What's the weather?", tools=[...])

Full Message Format

response = anyllm.chat(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is Python?"},
    ],
    model="openai/gpt-4",
    temperature=0.7,
    max_tokens=500,
)

Or use the system parameter shorthand:

response = anyllm.chat(
    "What is Python?",
    model="openai/gpt-4",
    system="You are a helpful assistant.",
)

Configuration

Environment Variables

Variable	Description
OPENAI_API_KEY	OpenAI API key
OPENAI_BASE_URL	Custom OpenAI-compatible API base URL
ANTHROPIC_API_KEY	Anthropic API key
OLLAMA_HOST	Ollama server URL (default: http://localhost:11434)
ANYLLM_DEFAULT_MODEL	Default model to use

Config File

Create ~/.anyllm/config.json:

{
  "default_model": "openai/gpt-4",
  "openai_api_key": "sk-...",
  "ollama_base_url": "http://localhost:11434"
}

Programmatic Configuration

import anyllm

# Set default model
anyllm.set_default("ollama/llama3")

# Check available providers
print(anyllm.available_providers())  # ['openai', 'ollama']

# List models
print(anyllm.list_models())
# {'openai': ['gpt-4', 'gpt-3.5-turbo', ...], 'ollama': ['llama3', ...]}

# Direct config access
config = anyllm.get_config()
config.set("openai_base_url", "http://localhost:1234/v1")

Response Object

response = anyllm.chat("Hello!", model="openai/gpt-4")

response.content        # "Hello! How can I help you?"
response.model          # "gpt-4"
response.usage          # Usage(prompt_tokens=10, completion_tokens=8, total_tokens=18)
response.raw_response   # Raw API response dict

Local-First / Edge AI

anyllm is designed with a local-first philosophy. When auto-detecting providers, it checks local options (Ollama, llama.cpp) before cloud APIs. This means your applications work offline by default when local models are available.

import anyllm

# See what local models are available
print(anyllm.list_local_models())
# {'ollama': ['llama3', 'mistral', ...]}

# Auto-detection prefers local -- this uses Ollama if running
response = anyllm.chat("Hello!")

# Explicitly use a local model
response = anyllm.chat("Hello!", model="ollama/llama3")

Error Handling

anyllm includes automatic retry with exponential backoff for transient errors:

# Retries up to 3 times by default
response = anyllm.chat("Hello!", model="openai/gpt-4", max_retries=3)

License

MIT License. See LICENSE for details.

Author

Viet-Anh Nguyen -- nrl.ai -- GitHub