Installing the MCP server #

lvkit's AI-agent surface is a local, stdio MCP server — the agent's runtime launches lvkit on your machine and it reads .vi files off your disk. This page gets that server in front of your agent, from a one-click install for a non-developer to a hand-written config file.

One-click for your app #

No terminal, no Python. Pick your app.

Claude Desktop #

Download the lvkit-<target>.mcpb bundle for your machine (darwin-arm64, darwin-x64, or win32-x64) from the GitHub Release or pragmatest.com, then either double-click it, drag it into Claude Desktop, or use Settings → Extensions → Install Extension… In the extension's settings, pick the folder containing your .vi files — lvkit reads VIs from there so you don't have to type full paths.

macOS and Windows only (Claude Desktop has no Linux build). The macOS binary is unsigned in v1 — if Gatekeeper blocks it, clear the quarantine flag on the downloaded file:

xattr -dr com.apple.quarantine lvkit-darwin-*.mcpb

VS Code #

Install LVKit from the VS Code Marketplace. The extension bundles a signed standalone binary and auto-registers the lvkit mcp server (VS Code ≥ 1.101), so agent mode gets the lvkit tools with no mcp.json, no Python, and no uv. See VS Code extension.

Claude Code (plugin) #

claude plugin marketplace add pragmatest-dev/claude-plugins
claude plugin install lvkit-<platform>@pragmatest

<platform> is one of mac-arm64, mac-intel, windows, linux — pick the one matching your machine. The plugin bundles the lvkit binary and lvkit's five workflow skills (the same ones setup installs), so it needs no Python, uv, or pip, and no lvkit on PATH.

Developer: configure your agent's config file #

Every client below points a command at lvkit and passes mcp as the argument (some clients spell that as args: ["mcp"], others fold it into the same invocation). That command is either of two equally valid things:

• uvx, which fetches lvkit from PyPI and runs it on demand — nothing to install first if you already have uv.
• The downloaded lvkit binary, as an absolute path — the standalone lvkit-mcp-<version>-<target>.zip from the GitHub Release, unzipped to lvkit/lvkit (lvkit\lvkit.exe on Windows). No Python, uv, or pip needed; useful for a client or machine that avoids package managers.

Pick whichever you already have. Both are shown for each client below.

Claude Code #

# uvx
claude mcp add lvkit -- uvx --from lvkit lvkit-mcp
 
# downloaded binary
claude mcp add lvkit -- /abs/path/to/lvkit mcp

VS Code #

.vscode/mcp.json:

{
  "servers": {
    "lvkit": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "lvkit", "lvkit-mcp"]
    }
  }
}

{
  "servers": {
    "lvkit": {
      "type": "stdio",
      "command": "/abs/path/to/lvkit",
      "args": ["mcp"]
    }
  }
}

Or skip the config file entirely — the VS Code extension registers the server for you.

GitHub Copilot CLI #

~/.copilot/mcp-config.json:

{
  "mcpServers": {
    "lvkit": {
      "type": "local",
      "command": "uvx",
      "args": ["--from", "lvkit", "lvkit-mcp"]
    }
  }
}

{
  "mcpServers": {
    "lvkit": {
      "type": "local",
      "command": "/abs/path/to/lvkit",
      "args": ["mcp"]
    }
  }
}

Copilot running inside VS Code instead uses the VS Code extension's zero-config registration — this file is for the standalone Copilot CLI.

Codex #

.codex/config.toml (shared by the Codex CLI, the Codex IDE extension, and the ChatGPT desktop app):

[mcp_servers.lvkit]
command = "uvx"
args = ["--from", "lvkit", "lvkit-mcp"]

[mcp_servers.lvkit]
command = "/abs/path/to/lvkit"
args = ["mcp"]

Codex and Copilot CLI have no one-click bundle — the config file above is their install path.

Cursor / any client using the standard mcpServers shape #

{
  "mcpServers": {
    "lvkit": {
      "command": "uvx",
      "args": ["--from", "lvkit", "lvkit-mcp"]
    }
  }
}

Alternatives #

Prefer a regular Python install over uvx or the standalone binary?

pip install lvkit      # or: uv tool install lvkit / pipx install lvkit

then point any client's command at the installed lvkit console script with args: ["mcp"] — the same shape as the downloaded-binary examples above, just without the absolute path. (lvkit mcp and the lvkit-mcp console script are the same server.)

Verify it works #

lvkit mcp --selftest   # initializes the server, lists tools, exits non-zero on failure

With uvx: uvx --from lvkit lvkit mcp --selftest (the bare lvkit-mcp script only runs the stdio server and ignores --selftest).

Notes #

• The claude.ai connector directory and the ChatGPT Apps directory host remote MCP servers. lvkit reads your local VI repo, so there is nothing for a hosted server to see — every path here is local by necessity.
• The workflow skills (lvkit, lvkit-describe, lvkit-query, lvkit-convert, lvkit-document, lvkit-review, lvkit-resolve) ship inside the Claude Code plugin bundle above; outside the plugin, install them separately with lvkit setup.

See also #

• mcp — the server's tools and a worked project-understanding demo.
• vscode-extension — the VS Code extension's view/diff/MCP surface in full.
• setup — install the workflow skills outside the Claude Code plugin.