Show HN: Allos – An open-source, LLM-agnostic agentic SDK for Python (github.com)

Allos is a new open-source, LLM-agnostic agentic SDK for Python that lets you build production-ready AI agents without vendor lock‑in. The MVP already supports OpenAI and Anthropic (Claude) with a unified Agent API and CLI, and roadmap items include Ollama/local models, Google Gemini, MCP support, web search/fetch tools, and more. Key built-ins: file ops, shell execution, session save/resume, fine‑grained tool permissions, automatic context-window checks, logging/error handling, and an extensible tool system so you can plug in custom tools or providers. Installation is simple (pip/uv), and the same AgentConfig interface can switch providers or models with no code rewrite. For ML practitioners the significance is practical: Allos formalizes a provider-agnostic orchestration layer so you can assign the best model for each task (e.g., GPT-4 for one job, Claude for another) and move between hosted and local runtimes more easily. Technical caveats for early adopters: no streaming, no async core, basic context compaction and error-recovery only, and web tools/providers beyond OpenAI/Anthropic are still on the roadmap. Overall, Allos offers a sensible, extensible foundation for multi‑provider agent development and experimentation while warning users about current MVP limitations around real-time responses and advanced context management.