lua-platform-mcp

Read-only MCP server for the Lua AI platform — exposes 5 tools that let any MCP-aware client (Cline, Cursor, Claude Code, Codex CLI, Continue, Zed, etc.) query Lua agent state mid-conversation.

Tools exposed

Tool	What it returns
list_agents	All agents accessible to the authenticated user ({id, name, orgId, orgName}[])
get_agent	Public-facing config for one agent by ID
list_primitive_versions	Versions of a skill / webhook / job / preprocessor / postprocessor / persona
get_deployment_status	What's live in production right now (composed from per-type version endpoints)
tail_logs	Structured log fetch (mirrors lua logs --json)

All tools are read-only — this server cannot push, deploy, or modify any state.

Install

Cline / any MCP client (recommended)

After cloning + building, register the server in your client's MCP config. For Cline, that's the cline_mcp_settings.json:

{
  "mcpServers": {
    "lua-platform": {
      "command": "node",
      "args": ["/absolute/path/to/lua-platform-mcp/dist/server.js"],
      "env": {
        "LUA_API_KEY": "lk_..."
      }
    }
  }
}

From source

git clone https://github.com/lua-ai-global/lua-platform-mcp
cd lua-platform-mcp
npm install
npm run build
# Now dist/server.js is the entry point

From npm (once published)

npx lua-platform-mcp   # one-shot
# or, in MCP config:
# "command": "npx", "args": ["-y", "lua-platform-mcp"]

Authentication

The server reads your Lua API key in this order:

1. LUA_API_KEY environment variable
2. ~/.lua-cli/credentials (written by lua auth configure)
3. .env file in CWD

Get a key from admin.heylua.ai (sign up + create an agent → Settings → API Keys), or use the Lua CLI:

npm install -g lua-cli
lua auth configure

Verify it works

After registering, restart your MCP client and ask: "What lua-platform tools do you have?"

You should see all 5 tools listed under the mcp__lua-platform__* prefix (Cline / Cursor / Claude Code) or as direct names (Codex / Continue).

Try: "List my Lua agents." → should return your agents, IDs, and orgs.

Architecture

• src/server.ts — MCP stdio bootstrap (uses @modelcontextprotocol/sdk)
• src/tools/*.mjs — one file per tool, each exporting { name, description, inputSchema, handler }
• src/auth.mjs — credential resolution (mirrors lua-cli's chain)
• src/api-client.mjs — typed HTTP wrapper around https://api.heylua.ai
• dist/server.js — esbuild-bundled single-file output (built via npm run build)

Why "read-only"?

Mutating operations (lua push, lua deploy, etc.) live in the Lua CLI directly, behind the §3.3 deploy-safety contract that requires explicit user confirmation. We deliberately don't expose them via MCP because MCP tool calls don't carry the human-confirmation step. Read tools are safe; write tools belong in the CLI.

If you want a richer "build agents inside your IDE" experience (slash commands, scaffolding, deploy gates), install one of the IDE-specific Lua plugins instead — they bundle this MCP server alongside the workflow:

• lua-ai-global/cursor-lua-agent-builder (Cursor)
• lua-ai-global/claude-code-lua-plugin (Anthropic Claude Code)
• (Codex CLI port coming soon)

License

MIT © Lua AI