AgentRL

[Aria-200]

hi,I'm very happy that you've open-sourced such an advanced framework,I want to use the agentrl framework to train models to improve their performance. Originally, I called tools through SSE. Under the agentrl framework, how should I import my tools to train the model?

[sailaoda]

Hi @Aria-200, thanks for your interest in our framework!
To integrate your custom tools with AgentRL for training, you can configure them in the AgentDock module. Here's the general approach:

1. Add your tools in the agentdock-node-explore node configuration (located at AgentCPM-Explore/AgentDock/agentdock-node-explore/)
2. Configure the MCP servers for your tools in the mcp_servers/ directory
3. Build the Docker image - once built, your tools will be available for the training framework to use
   For detailed setup instructions, please check:

• AgentDock README
• AgentRL Documentation
  If you have specific questions about tool configuration or encounter any issues during setup, feel free to share more details and we'll be happy to help!

[PrincessGray]

Hi @Aria-200, thanks for your interest in AgentRL.

With our latest commit, we now support the official MCP protocol for sampling. You simply need to set --mcp_manager_url to your stdio or SSE connection command (e.g. http://localhost:8088/mcp) to directly use your server for sampling or training. However, please note that you will need to manage concurrency and simultaneous connection limits on your end.