1.12.0
The framework for AI agents now supports OpenAI-compatible providers like DeepSeek and Ollama natively.
CrewAI Inc has launched version 1.12.0 of its popular open-source framework for building and orchestrating collaborative AI agents. This significant release, led by founder Joao Moura, introduces native support for a suite of OpenAI-compatible AI providers, including OpenRouter, DeepSeek, Ollama, vLLM, Cerebras, and Alibaba's Dashscope. This move decouples CrewAI from relying solely on LiteLLM for model routing, giving developers more direct control and flexibility. The update also debuts a formal 'agent skills' system, allowing developers to equip individual AI agents with specialized capabilities, and implements automatic root_scope for hierarchical memory isolation to prevent data leaks between different agent crews.
Beyond core agent features, v1.12.0 significantly improves international accessibility by adding a complete Modern Standard Arabic translation of all documentation, alongside Arabic support for changelogs and release tools. For system robustness, it introduces Qdrant Edge as a new storage backend for the memory system and resolves a batch of critical bugs, particularly within the human-in-the-loop (HITL) flow system where users can intervene in agent tasks. The release also includes extensive refactoring to simplify internal code, deduplicate sync/async task execution, and pin dependencies to stable versions, marking a mature step forward for the 47k-star GitHub project.
- Adds native support for OpenAI-compatible providers like OpenRouter, DeepSeek, and Ollama, reducing dependency on LiteLLM.
- Implements a new 'agent skills' system and automatic memory isolation (root_scope) for building more capable and secure multi-agent crews.
- Fully translates documentation into Arabic and fixes critical bugs in the human-in-the-loop flow and memory saving systems.
Why It Matters
This update makes building complex, multi-model AI agent systems more flexible, accessible to Arabic speakers, and production-ready.