nvim-mcp

Make Neovim feel like Cursor. nvim-mcp is a Model Context Protocol server that gives an agent — such as Claude Code — full control over the Neovim session it is running inside: buffers, windows, diagnostics, LSP language intelligence (definition, references, hover, rename, code actions), and, crucially, terminals.

When you launch an agent from inside a Neovim :terminal, the agent isn't just editing files in a vacuum — it's living inside your editor. nvim-mcp lets it act like it: open a split, read a buffer, check LSP diagnostics, spin up a terminal, run a command, and read the output back — the same things you'd do by hand.

Why not just claudecode.nvim?

Plugins like claudecode.nvim expose the file being edited to the agent. That is useful, but the agent still can't see or drive the rest of the editor — most notably it cannot open a terminal inside Neovim, run something, and read the result. nvim-mcp connects to Neovim's own RPC channel, so the agent gets the whole editor, terminals included.

Related MCP server: mcp-interactive-terminal

How it works

Neovim is a first-class RPC server. Whenever Neovim spawns a child process — a :terminal, jobstart(), system() — it exports the address of its own msgpack-RPC socket in the NVIM environment variable:

$ nvim                    # inside Neovim, open a terminal with :terminal
$ echo $NVIM
/run/user/1000/nvim.1234.0

Any client that connects to that socket can drive the parent editor. nvim-mcp:

1. reads NVIM (the default; this is set automatically inside a :terminal),

2. connects to that socket over msgpack-RPC, and

3. exposes a focused set of MCP tools that the agent can call.

┌──────────────────────────── Neovim ────────────────────────────┐
│  your code buffers, windows, LSP diagnostics                    │
│                                                                 │
│  ┌─ :terminal ──────────────────────────────────────────────┐  │
│  │  $ claude            ($NVIM is set in here)               │  │
│  │     └── nvim-mcp ──msgpack-RPC──┐                         │  │
│  └─────────────────────────────────┼────────────────────────┘  │
│        controls buffers/terminals  │                            │
│  ◄─────────────────────────────────┘                            │
└─────────────────────────────────────────────────────────────────┘

Install

npm install -g nvim-mcp     # or run on demand with: npx -y nvim-mcp

From source:

git clone https://github.com/shooooooooo/nvim-mcp
cd nvim-mcp
npm install
npm run build

Use with Claude Code

Register the server so Claude Code can launch it. The simplest way is a project-level .mcp.json (see examples/mcp.json):

{
  "mcpServers": {
    "nvim": { "command": "npx", "args": ["-y", "nvim-mcp"] }
  }
}

or globally:

claude mcp add nvim -- npx -y nvim-mcp

Then start Claude Code from inside Neovim so it inherits $NVIM:

:terminal claude

That's it — the agent can now call the nvim_* tools and they act on the editor it's running inside. Ask it things like "open a terminal and run the tests, then fix whatever fails."

Tools

LSP tools — the editor's language intelligence

These let the agent use the language servers Neovim already has running, instead of re-deriving semantics from raw text. Positions are 1-based line and 1-based column.

Targeting a specific Neovim

By default the target is discovered from the environment, highest priority first:

1. --socket <addr> / --address <addr> CLI flag

2. NVIM_MCP_SOCKET

3. NVIM (set automatically inside a Neovim :terminal)

4. NVIM_LISTEN_ADDRESS (legacy)

<addr> may be a unix-socket path / named pipe, or a host:port TCP address (start Neovim with nvim --listen 127.0.0.1:6789 for the latter).

Testing

The test suite is real, not mocked: every integration test spins up an actual headless Neovim, points NVIM at it, and drives it exactly as the agent would.

npm test

What's covered:

• test/nvim.test.ts — the controller against a live Neovim: address resolution, session info, Lua/eval/Ex commands, buffer read/write, windows, diagnostics.

• test/terminal.test.ts — the headline scenario: open a terminal, run echo hello world, read it back; plus non-zero exit codes and interactive send/read.

• test/lsp.test.ts — the LSP tools driven against a deterministic fake language server (test/helpers/fake-lsp.mjs), covering hover, definition, references, document symbols, rename (apply + preview), code actions, and formatting — all offline, no real language server or network required.

• test/mcp.test.ts — the full MCP stack: an MCP client launches the server as a subprocess with NVIM set, lists the tools, and calls nvim_run_in_terminal to run echo hello world end-to-end over JSON-RPC.

The reusable harness in test/helpers/nvim.ts starts a pristine headless Neovim (-u NONE -i NONE -n) on a fresh socket and tears it down cleanly, so you can build more scenarios on top of it. CI (.github/workflows/ci.yml) runs the whole thing against multiple Node and Neovim versions.

Real end-to-end with the live agent

npm test deliberately avoids needing an API key — it drives the MCP server directly, which is the exact code path the agent uses. To verify the literal "Claude Code inside Neovim" loop with the real LLM, see scripts/e2e-claude-code.sh: it starts Neovim, launches Claude Code in an in-editor terminal, and asks it to run echo hello world in a Neovim terminal and read it back. That script needs the claude CLI authenticated and is kept out of the automated suite for that reason.

Development

npm install
npm run dev          # run the server from TypeScript (tsx)
npm run typecheck
npm run build
npm test

Releasing

Publishing to npm is automated by .github/workflows/release.yml, which runs on any pushed v* tag (or manually from the Actions tab). One-time setup: add an NPM_TOKEN repository secret (an npm automation token with publish rights). Then cut a release:

npm version patch     # or minor / major — bumps package.json and tags it
git push --follow-tags

The workflow typechecks, builds, and runs npm publish --provenance, so the published package always ships a fresh dist/.

License

MIT — see LICENSE.