Llmswap: Avoid LLM vendor lock-in – 10 providers with top LMArena models (github.com)

llmswap v5.1.0 launches a workspace-centric SDK and CLI that promises to eliminate LLM vendor lock‑in and turn any provider into a persistent, project-aware AI mentor. The release adds per-project “brains” (auto-created .llmswap with context.md, learnings.md, decisions.md), automated learning journals, architecture decision logs, six teaching personas, and workspace auto-switching so the assistant remembers codebases, past learnings, and decisions across sessions. It supports ten providers (OpenAI, Anthropic, Google Gemini, Cohere, Perplexity, IBM watsonx, Groq, Ollama, xAI Grok, Sarvam AI) and uses a pass-through architecture so new models can be used immediately by name (e.g., client = LLMClient(provider="openai", model="gpt-5") or llmswap chat --provider openai --model gpt-5). Technically significant for ML teams and developers, llmswap combines multi‑provider flexibility with production validation: defaults are chosen and updated weekly from LMArena top models (e.g., claude-sonnet-4-5, grok-4, gemini-2.0-flash-exp, gpt-4o-mini) and all integrations are battle‑tested with real API calls. It offers cost optimizations via caching (claiming 50–90% savings), a single code/CLI surface for switching providers, and RAG-friendly tooling for project-specific context. The net result is faster experimentation with new models, simpler multi‑provider deployments, and persistent, teachable assistants that reduce repetitive context setup and support long-term developer learning and architecture traceability.