llmgt 0.2.0

Plug-and-play LLM connector via YAML config

llmgate

Plug-and-play LLM connector via YAML config. One interface, 21 providers, zero bloat.

Why llmgate?

You've probably seen LiteLLM. It's great — if you want a proxy server, Redis, PostgreSQL, a dashboard, and 50+ transitive dependencies. If you just want to call an LLM from Python without installing a framework, there's nothing lightweight out there.

llmgate is the opposite: pip install llmgt pulls in exactly two dependencies (httpx + pyyaml). Drop a YAML file in your project, set your API key, and call any model. No proxy server, no database, no SDK lock-in — just a Python library that reads a config and makes HTTP calls. Swap providers by changing one line in your YAML.

	llmgate	LiteLLM
Install size	~2 MB	~200 MB+
Dependencies	2 (httpx, pyyaml)	50+
Architecture	Library (import it)	Proxy server
Provider swap	Change 1 line in YAML	Change code
Latency overhead	~0 (direct HTTP)	Proxy hop + DB logging

Note: The PyPI package is llmgt (pip install llmgt), but the import is llmgate.

Install

pip install llmgt

Optional extras:

pip install llmgt[aws]    # AWS Bedrock (boto3)
pip install llmgt[gcp]    # Google Vertex AI (google-auth)
pip install llmgt[dev]    # pytest + dev tools

Quickstart

1. Create llmgate.yaml in your project:

provider: anthropic
model: claude-sonnet-4-20250514
api_key: ${ANTHROPIC_API_KEY}
temperature: 0.7
max_tokens: 1024

2. Use it:

from llmgate import LLMGate

gate = LLMGate()
response = gate.chat("Explain transformers in one sentence")
print(response.text)
print(response.tokens_used)

# Streaming
for chunk in gate.stream("Write a haiku"):
    print(chunk, end="", flush=True)

System Prompts & Multi-Turn

For simple prompts use chat(). For system prompts or conversation history, use chat_messages() with the full messages list:

response = gate.chat_messages([
    {"role": "system", "content": "You are a helpful coding assistant. Be concise."},
    {"role": "user", "content": "What's a closure?"},
])
print(response.text)

Multi-Turn Conversations

Build up conversation history and pass it in:

messages = [
    {"role": "system", "content": "You are a math tutor."},
    {"role": "user", "content": "What's a derivative?"},
]

response = gate.chat_messages(messages)
print(response.text)

# Continue the conversation
messages.append({"role": "assistant", "content": response.text})
messages.append({"role": "user", "content": "Can you give me an example?"})

response = gate.chat_messages(messages)
print(response.text)

Streaming works with full message lists too:

for chunk in gate.stream_messages(messages):
    print(chunk, end="", flush=True)

Multi-Profile Config

active_profile: smart

defaults:
  temperature: 0.7
  max_tokens: 1024

profiles:
  smart:
    provider: anthropic
    model: claude-sonnet-4-20250514
    api_key: ${ANTHROPIC_API_KEY}

  fast:
    provider: groq
    model: llama-3.1-8b-instant
    api_key: ${GROQ_API_KEY}

  cheap:
    provider: deepseek
    model: deepseek-chat
    api_key: ${DEEPSEEK_API_KEY}

  local:
    provider: ollama
    model: llama3.2

Hot-swap profiles at runtime:

gate = LLMGate()                          # uses "smart" profile
gate.switch("fast")                       # swap to Groq
response = gate.chat("Hello", temperature=0.2)  # call-time overrides

Loading API Keys from .env

llmgate resolves ${ENV_VAR} from os.environ. To load keys from a .env file, use python-dotenv:

pip install python-dotenv

from dotenv import load_dotenv
load_dotenv()  # loads .env into os.environ

from llmgate import LLMGate
gate = LLMGate()  # now ${ANTHROPIC_API_KEY} etc. will resolve

Or use a .env file:

ANTHROPIC_API_KEY=sk-ant-...
GROQ_API_KEY=gsk_...
OPENAI_API_KEY=sk-...

See .env.example in the repo for all supported variables.

Environment Variable Interpolation

Any string value in the YAML can use ${ENV_VAR} syntax — not just api_key:

api_key: ${MY_API_KEY}
base_url: ${CUSTOM_ENDPOINT}
nested:
  deep:
    value: ${SOME_SECRET}

Variables are resolved from os.environ at load time. Missing vars resolve to empty string.

Supported Providers

Provider	Example Models	Env Var	Streaming	Notes
openai	gpt-4o, gpt-4-turbo	OPENAI_API_KEY	✅
anthropic	claude-sonnet-4-20250514, claude-opus-4-20250514	ANTHROPIC_API_KEY	✅
gemini	gemini-1.5-pro, gemini-1.5-flash	GEMINI_API_KEY	✅
cohere	command-r-plus, command-r	COHERE_API_KEY	✅
groq	llama-3.1-8b-instant, mixtral-8x7b	GROQ_API_KEY	✅	OpenAI-compatible
mistral	mistral-large, mistral-small	MISTRAL_API_KEY	✅	OpenAI-compatible
openrouter	meta-llama/llama-3.1-70b-instruct	OPENROUTER_API_KEY	✅	OpenAI-compatible
together	meta-llama/Llama-3-70b-chat-hf	TOGETHER_API_KEY	✅	OpenAI-compatible
fireworks	accounts/fireworks/models/llama-v3-70b	FIREWORKS_API_KEY	✅	OpenAI-compatible
perplexity	llama-3.1-sonar-large-128k	PERPLEXITY_API_KEY	✅	OpenAI-compatible
deepseek	deepseek-chat, deepseek-coder	DEEPSEEK_API_KEY	✅	OpenAI-compatible
xai	grok-2, grok-beta	XAI_API_KEY	✅	OpenAI-compatible
ai21	jamba-1.5-large, jamba-1.5-mini	AI21_API_KEY	✅	OpenAI-compatible
azure_openai	gpt-4o (via deployment)	AZURE_OPENAI_API_KEY	✅	See Azure setup
bedrock	anthropic.claude-3, amazon.titan	AWS credentials	❌	See Bedrock setup
vertexai	gemini-1.5-pro (via Vertex)	GCP ADC	✅	See Vertex setup
huggingface	mistralai/Mixtral-8x7B-Instruct-v0.1	HUGGINGFACE_API_KEY	❌	Auto-detects chat models
replicate	meta/llama-2-70b-chat	REPLICATE_API_KEY	❌	Polling-based
nlpcloud	chatdolphin, finetuned-llama-3	NLPCLOUD_API_KEY	❌
ollama	llama3.2, mistral, codellama	none	✅	Local
lmstudio	any GGUF model	none	✅	Local, OpenAI-compatible

Providers marked ❌ for streaming will return the full response as a single chunk when you call stream().

Error Handling

llmgate raises standard exceptions you can catch:

import httpx
from llmgate import LLMGate

gate = LLMGate()

try:
    response = gate.chat("Hello")
except FileNotFoundError:
    # llmgate.yaml not found
    print("Create a llmgate.yaml config file first")
except ValueError as e:
    # Bad config: unknown provider, missing profile, missing 'provider' field
    print(f"Config error: {e}")
except httpx.HTTPStatusError as e:
    # API returned an error (401 unauthorized, 429 rate limited, 500 server error, etc.)
    print(f"API error {e.response.status_code}: {e.response.text}")
except httpx.ConnectError:
    # Can't reach the API (network issue, wrong base_url, Ollama not running)
    print("Connection failed — check your network or base_url")
except httpx.TimeoutException:
    # Request took longer than 60 seconds
    print("Request timed out")
except ImportError as e:
    # Missing optional dependency (boto3 for Bedrock, google-auth for Vertex)
    print(f"Missing dependency: {e}")

All API errors come through as httpx.HTTPStatusError with the full response body available at e.response.text — useful for debugging rate limits, auth issues, or quota problems.

Azure OpenAI Setup

profiles:
  azure:
    provider: azure_openai
    model: gpt-4o
    resource_name: my-azure-resource
    deployment_name: my-gpt4o-deployment
    api_version: "2024-02-01"
    api_key: ${AZURE_OPENAI_API_KEY}

AWS Bedrock Setup

pip install llmgt[aws]

profiles:
  aws:
    provider: bedrock
    model: anthropic.claude-3-sonnet-20240229-v1:0  # or amazon.titan-*, meta.*
    region: us-east-1

Requires AWS credentials configured via ~/.aws/credentials, env vars, or IAM role. Supports Anthropic Claude, Amazon Titan, and Meta Llama model families on Bedrock — detected automatically by model ID prefix.

Google Vertex AI Setup

pip install llmgt[gcp]

profiles:
  gcp:
    provider: vertexai
    model: gemini-1.5-pro
    project_id: my-gcp-project
    region: us-central1

Uses Google Application Default Credentials. Run gcloud auth application-default login or set GOOGLE_APPLICATION_CREDENTIALS.

LLMResponse

response = gate.chat("Hello")
response.text           # str — the generated text
response.model          # str — model name
response.provider       # str — provider name
response.tokens_used    # int | None — total tokens
response.finish_reason  # str | None — stop reason
response.raw            # dict — full API response

Async Support

Not yet — llmgate v0.1 is sync-only (httpx sync client). Async via httpx.AsyncClient is planned for v0.2. If this is blocking you, open an issue.

Contributing

git clone https://github.com/kesiee/llmgate.git
cd llmgate
pip install -e ".[dev]"
pytest

The codebase is intentionally simple. Provider files live in llmgate/providers/. OpenAI-compatible providers inherit from OpenAIProvider and only override BASE_URL + headers. Custom providers implement send() and stream() directly.

To add a new provider:

1. Create llmgate/providers/yourprovider.py — inherit from BaseProvider (or OpenAIProvider if compatible)
2. Add it to PROVIDER_REGISTRY in llmgate/gate.py
3. Add a test and update this README

License

MIT