MCP servers

litecode is an MCP client. It starts MCP servers over stdio and gives their tools to the model. It reads the standard mcpServers config format that most MCP clients use.

Add a server#

Command line:

litecode mcp add fs npx -y @modelcontextprotocol/server-filesystem .
litecode mcp add git uvx mcp-server-git
litecode mcp                 # list servers
litecode mcp --tools         # start them and list their tools
litecode mcp rm fs

Or edit a config file:

{
  "mcpServers": {
    "fs": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]},
    "notes": {"command": "python", "args": ["notes_server.py"], "env": {"NOTES_DIR": "D:/notes"}},
    "off": {"command": "x", "disabled": true}
  }
}

Config files (later wins on a name clash):

1. <data_dir>/mcp.json (user, written by litecode mcp add)
2. <project>/.mcp.json (project, shared with other MCP clients)
3. <project>/.litecode/mcp.json

How the model calls a tool#

Tools appear as mcp__<server>__<tool>. The model uses the normal JSON format:

{"tool": "mcp__fs__read_file", "args": {"path": "README.md"}}

Each MCP call asks for approval in ask mode. In readonly mode MCP calls are refused.

Limits#

• Transport: stdio only. HTTP/SSE servers are not supported yet.
• Tools only. MCP resources and prompts are not used yet.
• litecode shows the model at most 20 MCP tools. Small models choose badly from long lists. Add only what you need.
• A server that fails to start is skipped. litecode prints the error and continues.
• Most MCP servers need Node.js (npx) or uv (uvx). Install those first. Offline use needs the server installed locally.

Source: docs/MCP.md