CowAgent 2.0.9 lands three changes that matter if you are building on top of it: a unified model management UI, native MCP protocol support, and persistent browser sessions for login-gated sites.
Model management finally makes sense. The web console now has a dedicated Models page. It organizes configuration by provider and capability, covering chat, image, voice, embedding, and search. The key improvement is unified provider config: you enter an API key and base URL once at the top, and every downstream capability picks it up automatically. No more entering the same key in multiple places. Image understanding and image generation can each choose their own provider and model independently, or fall back to the main chat model. STT and TTS are configured separately, with Qwen and GLM ASR and TTS models now supported. Embedding models for memory and knowledge-base retrieval support OpenAI, Qwen, Doubao, GLM, and others. If you switch embedding models, run /memory rebuild-index to rebuild the index. Web search now supports Bocha, Baidu, GLM, and more, with an auto mode that lets the agent combine multiple sources for research tasks. Full details at the Models Overview docs.
MCP support turns CowAgent into an open tool platform. The agent now implements the Model Context Protocol natively, with zero extra dependencies. Both stdio and sse transports are supported. Any MCP-compatible service plugs in directly as a tool, which means the fixed built-in tool set is no longer the ceiling. Configuration follows the mcpServers format used by Claude Desktop and Cursor, and ~/cow/mcp.json is loaded first. See the MCP Tools docs for setup.
Browser sessions now persist. For sites that require login or use anti-bot protection, the browser tool can maintain login state across runs. This makes the agent usable for workflows that previously broke on session expiry.
One security note before you upgrade. This release changes the web console to bind to 127.0.0.1 by default. If you are running CowAgent on a server and need public access, set web_host to 0.0.0.0 in config.json. The release notes strongly recommend also setting web_password to enable authentication at that point. Check the Manual Installation guide before upgrading on any server.
If you are building an agent workflow today, the MCP integration is the most actionable change. Pick an MCP-compatible service you already use, drop the config into ~/cow/mcp.json, and you have a new tool available without writing any adapter code.