Six tools. One shell. AI agents get Nushell—without Bash baggage.

Most MCP servers ship 20+ tools. One for listing files, another for searching, three for JSON parsing. The agent spends half its time figuring out which tool does what.

Nushell already has 400+ commands that handle pipelines, structured data, and system operations. nu-mcp exposes those through six tools:

Result: Fewer tools to maintain, better performance, and agents that understand what they're doing.

cargo install nu-mcp

Or build from source:

git clone https://github.com/Xpos587/nu-mcp
cd nu-mcp
cargo build --release

Binary ends up at target/release/nu-mcp.

Claude Code:

claude mcp add nu-mcp \
  --transport stdio \
  --scope user \
  --env SEARXNG_URL=http://127.0.0.1:8888 \
  --env APPLY_API_KEY=ollama \
  --env APPLY_API_URL=http://localhost:11434/v1 \
  --env APPLY_MODEL=morph-v3-fast \
  -- /path/to/nu-mcp

Codex:

codex mcp add nu-mcp \
  --env SEARXNG_URL=http://127.0.0.1:8888 \
  --env APPLY_API_KEY=ollama \
  --env APPLY_API_URL=http://localhost:11434/v1 \
  --env APPLY_MODEL=morph-v3-fast \
  -- /path/to/nu-mcp

Run Nushell commands.

Blocking:

command: "ls src | where size > 1kb"

Returns:

Exit code: 0
Time: 23ms

file1.rs
file2.rs

Background:

command: "cargo watch"
background: true

Returns job ID. Use nu.output to get results.

Options:

Get output from background process.

id: "job_abc123"

Returns:

ID: job_abc123
Status: running
Running for: 15s
Exit code: (running)

Finished dev [unoptimized + debuginfo] target(s) in 0.52s

Stop background process.

id: "job_abc123"

Returns:

ID: job_abc123
Status: killed
Command: cargo watch

Edit files with Fast Apply markers.

path: "/path/to/file.rs"
instructions: "Add error handling"
code_edit: |
  // ... existing code ...

  fn main() -> Result<()> {
      do_work()?;
      Ok(())
  }

  // ... existing code ...

Returns:

Path: /path/to/file.rs
Status: applied
Message: Code edit applied to /path/to/file.rs

Why this matters: Closed Fast Apply services lock you into their infrastructure. Local models (Ollama, vLLM) give you privacy and control.

Search web, packages, and code via SearXNG.

query: "tokio"
category: "cargo"

Returns:

Query: "tokio" | Category: cargo | Found: 150 | Showing: 3

[1] tokio - Rust asynchronous runtime
    URL: https://crates.io/crates/tokio
    Engine: crates.io
    Content: An event-driven, non-blocking I/O platform...

[2] async-std - Async standard library
    URL: https://crates.io/crates/async-std
    Engine: crates.io

Categories: general, cargo, packages, it, repos, code, skills, science, news, videos, images, books, files

Options:

Fetch web content. HTML → Markdown automatic.

url: "https://example.com"

Returns:

URL: https://example.com
Status: 200
Content-Type: text/html
Format: markdown

# Example Domain
...

Options:

Syntax notes:

• Use ; for chaining, not &&
• Pipes work with structured data
• Variables: $name, $env.PATH
• External commands: ^git status

src/
├── main.rs    — MCP server, tool handlers
├── exec.rs    — Command execution, background jobs, Fast Apply, search, fetch
└── state.rs   — CWD tracking, process registry

Design choices:

1. Stateful sessions — Working directory persists between commands
2. Active pipe draining — Reads stdout/stderr concurrently to prevent hangs
3. Kill-on-timeout — No zombie processes
4. Markdown stripping — nu.apply removes code fences before writing
5. Plain text output — All tools return human-readable text, not JSON

cargo build --release
RUST_LOG=debug cargo run
cargo fmt
cargo clippy
cargo test

MIT

• Nushell — Modern shell for structured data
• SearXNG — Privacy-respecting metasearch engine
• MorphLLM — Original Fast Apply implementation
• Model Context Protocol — AI tool integration standard
• Claude Code — Claude CLI with MCP
• Codex — AI terminal with MCP