Support iterative discovery of MCP servers and tools

[chris-boson]

Read tool definitions on-demand from the filesystem, rather than reading them all up-front.

The agent can then write code to interact with MCP servers / tools.

https://www.anthropic.com/engineering/code-execution-with-mcp

[eyurtsev]

There was another request for this feature already. We'll likely add it at some point.

• Users can implement this right now using middleware
• Unclear whether it needs to be coupled to the file system? The main idea is iterative discovery

[chris-boson]

There is also this library that already has an integration with langchain (and deep agents by extension) https://github.com/maxvaega/skillkit

I think one thing that is missing is a tight coupling to file system and code execution

[Chen-zexi]

Hi @chris-boson , here is an implementation of the code-execution-with-mcp based on DeepAgents: https://github.com/Chen-zexi/open-ptc-agent