This document explains how MCP works in DeepSeek Code today, how to configure it, and what is planned.
DeepSeek Code currently provides:
- MCP server configuration in
.deepseek-code.json - MCP visibility in the TUI (
/mcp) - MCP runtime toggling in the TUI (
/mcp enable <name>,/mcp disable <name>) - MCP status in the shortcut panel and status bar
DeepSeek Code does not yet execute MCP protocol calls against servers in the agent loop.
In short: configuration + UX are implemented, full tool-bridging is the next step.
Add mcpServers in .deepseek-code.json:
{
"provider": "deepseek",
"model": "deepseek-chat",
"apiKey": "env:DEEPSEEK_API_KEY",
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."],
"enabled": true
},
"git": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-git"],
"enabled": false
}
}
}command: executable to runargs: argument arraycwd: optional working directory for the server processenv: optional environment map ("env:NAME"is supported)enabled: optional boolean, defaults to enabled
/mcp(or/mcp list) — list configured servers/mcp enable <name>— enable one server/mcp disable <name>— disable one server
These commands update in-memory session state.
- Add an
MCPManagerthat starts/stops stdio MCP processes. - Pull tool schemas from MCP servers.
- Convert MCP tools to AI SDK tools at runtime.
- Route tool calls through the manager and stream results.
- Persist enabled/disabled state to session file.
This keeps DeepSeek Code modular and compatible with any MCP server ecosystem.