Last year I built an AI agent with standard tools for things like file access. However, I noticed that I wanted to give the user a richer experience, I wanted to show clear diffs when the LLM proposed changes, I wanted to allow the user to view the file system as a whole without having to switch to another tool or cobble together an alternative view that matched the tool configuration.
However, this forced me to go outside the standard specifications for such tools and build a custom solution for my specific agent.
In the MCP4H proposal I have tried to bring that UI power to a standard MCP server by layering some additional (optional) features on top of it.
It is inspired by MCP Apps and uses some MCP v2 features to offer complex visualization on top of the standard tools aimed at the LLM.
Note that it is entirely opt-in and the MCP file server I added as an example can be used as a standard MCP server as well.
In addition to the UI components, I have added a standardized "dryrun" proposition using the new "_meta" fields. This can use the visualization to better inform the user of what he might be agreeing to. I've also added a permission model example that tries to balance ease of use with detailed permission management.
Show HN: MCP4H – A human-centric extension for the Model Context Protocol · HackerTrans
However, this forced me to go outside the standard specifications for such tools and build a custom solution for my specific agent.
In the MCP4H proposal I have tried to bring that UI power to a standard MCP server by layering some additional (optional) features on top of it.
It is inspired by MCP Apps and uses some MCP v2 features to offer complex visualization on top of the standard tools aimed at the LLM.
Note that it is entirely opt-in and the MCP file server I added as an example can be used as a standard MCP server as well.
In addition to the UI components, I have added a standardized "dryrun" proposition using the new "_meta" fields. This can use the visualization to better inform the user of what he might be agreeing to. I've also added a permission model example that tries to balance ease of use with detailed permission management.