Vision

Make nushell MCP fully self-sufficient so that AI coding assistants (like Claude Code) can operate entirely through nushell — no separate file tools needed.

Current State

Today, Claude Code uses nushell MCP for shell evaluation but still relies on its own built-in tools for:

• Reading files (Read tool)
• Writing files (Write tool)
• Editing files (Edit tool)
• Searching (Glob, Grep tools)

Goal

Nushell already has all these capabilities natively:

• open / save for file I/O
• str replace for editing
• glob / ls **/* for finding files
• Built-in string manipulation

If we expose the right primitives and optimize for AI agent usage, nushell MCP could be the only tool an AI needs.

Why This Matters: Remote Sessions

The killer feature: SSH-based remote coding.

If everything goes through nushell MCP over stdio:

1. SSH to a remote machine
2. Start nu --mcp
3. Connect Claude Code to that stdio stream
4. Everything works remotely — file reads, writes, searches, all over the network

No need for separate file sync, no sshfs, no remote VS Code servers. Just nushell.

Potential Enhancements

1. Transport options — Currently only stdio, could add streamable HTTP (rmcp already supports transport-streamable-http-server)
2. Optimized file operations — Perhaps dedicated MCP tools for common patterns (read file with line numbers, atomic edit, etc.)
3. Session state — The $history variable (feat(mcp): add $history variable and structured response format #17132) is a good start for maintaining context

Related

• feat(mcp): add $history variable and structured response format #17132 — $history variable for structured responses
• MCP: Piping non-nushell commands hangs forever (e.g., echo "hi" | cat) #17122 — Piping non-nushell commands
• NUON: Support raw strings (SNUON) for token-efficient LLM serialization #17124 — SNUON for token-efficient serialization