Example SRE workflow using different providers

I Built an Open-Source SDK to Use Any LLM Provider. Get Started in 2 Minutes.

Avatar photoByAdil

Hey Dev Community!

For the last few weeks, I’ve been heads-down building a tool to solve a problem that’s been driving me crazy: vendor lock-in when building AI agents.

I love the power of agents, but I hated that choosing a framework meant getting stuck with one provider. Want to switch from GPT-5 to Claude 4? That meant a painful rewrite.

So, I built Allos: an open-source, MIT-licensed agentic SDK for Python. The philosophy is simple: write your agent logic once, run it with any LLM provider.

Show Me the Magic: 2-Minute Demo

Before I explain more, just watch what it can do. Here’s a 2-minute video where Allos builds a complete, multi-file FastAPI application from a single prompt.

Get Started in 2 Minutes

You can replicate that experience right now.

1. Install the SDK (with all providers): 

uv pip install "allos-agent-sdk[all]"

2. Set your API key: 

export OPENAI_API_KEY="your_api_key_here"

3. Run your first agent! 

allos "Create a simple FastAPI app in main.py and tell me how to run it."

That’s it. The agent will plan the steps, show you the code it’s about to write, ask for your permission, and get it done.

The Core Promise: True Provider Freedom

This is the main event. You can start a task with one provider and finish it with another, using the exact same code and session context. 

# Start with OpenAI
allos --provider openai "Refactor this code for readability."

# Switch to Anthropic for deeper reasoning
allos --provider anthropic --session refactor.json "Continue the analysis."

Example SRE workflow using different providers

Why is this simpler? Less Boilerplate.

Frameworks should get out of your way. Allos is designed to be minimal and intuitive.

Langchain basic implementation

Similar implementation with Allos

What’s Included in the Box?

Allos isn’t just a wrapper; it’s a full toolkit for building agents.

  • A Polished CLI: The allos command is your main entry point. It supports single-shot tasks, a full interactive mode (-i), and session management (-s).

Allos CLI

  • Secure, Built-in Tools:

    • Filesystem: read_file, write_file, edit_file, list_directory
    • Execution: shell_exec (with timeouts and a dangerous command blocklist)
    • A human-in-the-loop permission system means you’re always in control.

  • Extensible by Design: Creating your own tool is as simple as writing a Python class with a @tool decorator. 

    @tool
class DatabaseQueryTool(BaseTool):
    name = "query_database"
    # ... your implementation

The Roadmap: Building an Open Ecosystem

This MVP is just the beginning. The public roadmap is focused on what developers need next:

  • 🔌 Ollama Support: First-class support for running local models is our next major priority.
  • 🌐 Web Tools: Built-in web_search and web_fetch tools + MCP!.
  • 🧠 More Providers: Google Gemini, Cohere, and more.
  • 🤖 More Agentic Frameworks: smolagents, Pydantic AI, and more.

Join the Community!

This project is built in the open, for the community. If you believe developers should have the freedom to choose their tools, I’d love your support.

We have a Stargazers Hall of Fame in our README to thank our earliest supporters!

I’m incredibly proud of this first release and can’t wait to see what you build with it. Let me know what you think in the comments

Similar Posts