Model Context Protocol¶

Experimental

MCP support is experimental and may change in future releases.

The Model Context Protocol (MCP) is an open standard that enables AI assistants to connect with external tools and data sources.

Limitations¶

stdio-based MCPs only - Network-based transports are not supported
Tools only - Resources and prompts are not supported

Adding an MCP Server¶

Use /mcp-add to register an MCP server:

/mcp-add <name> [<env>...] <cmd...>

<name> - A unique identifier for the MCP server (used to invoke its tools)
<env> - Optional environment variables in KEY=VALUE format
<cmd> - The command to start the MCP server

Examples¶

Add the Git MCP server:

/mcp-add ddg uvx -q duckduckgo-mcp-server

When an MCP server is added, the tool usage schema is added to the conversation. This may consume a substantial number of tokens.

Ask LLM to call an MCP¶

MCP tools can be called using the !hai-tool:

[1] !hai search "prusa core one indx release"

↓↓↓


- /mcp_ddg search {"query": "prusa core one indx release", "max_results": 10}


⚙ ⚙ ⚙

Pushed 1 command(s) into queue

---

!hai-tool[0]: /mcp_ddg search {"query": "prusa core one indx release", "max_results": 10}
[03/02/26 17:03:21] INFO     Processing request of type            server.py:720
                             CallToolRequest                                    
[03/02/26 17:03:22] INFO     HTTP Request: POST                  _client.py:1740
                             https://html.duckduckgo.com/html                   
                             "HTTP/1.1 200 OK"                                  
Found 10 search results:

1. Introducing the INDX! Fast and affordable 8-material printing ...

...

Note that the LLM called the MCP tool using the regular command /mcp_ddg, which is discussed in the next section.

Call MCP Tools Manually¶

Once registered, call MCP-tools using the dynamically created command:

/mcp_<name> <tool_name> <json_arg>

<name> - The name you assigned when adding the MCP server
<tool_name> - The name of the tool to invoke
<json_arg> - A JSON object containing the tool's arguments