dsh-mcp-json

You have MCP servers configured and working in Claude Code, Kiro, OpenCode, or Codex, yet switching to DeepSeek Harness means re-editing the plugin tree and retyping every command, args, and env block from scratch. dsh-mcp-json automates that hand-off: it scans ten user-level and project-level paths where those tools keep their configs, mounts every MCP server it finds onto the Harness MCP client, and lets you keep editing plain JSON instead of wrestling with a plugin profile.

Dialect differences add another layer of friction. Claude uses mcpServers with command and args; OpenCode calls it mcp and wraps argv in a single array; Codex writes TOML under mcp_servers. The plugin ships a built-in translator that normalises field names, transport tags (stdio vs http), and switch semantics so you never maintain one copy per tool. Files may carry line comments, block comments, and trailing commas, because that is what real editor output looks like.

Live reload works per server, not by remounting the whole set. A server whose resolved configuration is unchanged keeps its connection and registered tools while only the added, removed, or edited one is reconciled. Editor writes arrive as several file events, which a 150 ms debounce window coalesces before the files are re-read, avoiding half-written documents. If you would rather own your configuration outright, a one-shot import command translates every borrowed entry into ~/.dsh/mcp.json; flipping borrow to false then locks discovery to your own file so nothing rewrites it on the next start.

This plugin is for developers who juggle two or three AI coding tools, already have MCP servers configured in at least one of them, and want those servers to ride along in DeepSeek Harness without re-typing a single field. It is equally useful in mixed-tool teams where one person ships a Kiro config and another ships an OpenCode config, and the shared runtime needs to see both without anyone maintaining a second, hand-synced copy.