mcp-agentify

AI-powered MCP gateway that discovers tools from backend MCP servers and routes each natural-language request to one backend tool.

Experimental. The npm package is @steipete/mcp-agentify. The unscoped mcp-agentify package belongs to an unrelated project.

Requirements

• Node.js 20.19 or newer
• An OpenAI API key
• At least one stdio MCP backend

Install

npm install --global @steipete/mcp-agentify

The installed command remains mcp-agentify.

Configure

Create mcp-agentify.json:

{
  "openaiModel": "gpt-4.1-mini",
  "frontendPort": 3030,
  "agents": ["openai/gpt-4.1-mini"],
  "backends": [
    {
      "id": "filesystem",
      "displayName": "Workspace files",
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem@2026.1.14",
        "/absolute/path/to/allowed/files"
      ]
    },
    {
      "id": "browserbase",
      "displayName": "Browserbase",
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@browserbasehq/mcp@3.0.0"],
      "inheritEnv": [
        "BROWSERBASE_API_KEY",
        "BROWSERBASE_PROJECT_ID",
        "GEMINI_API_KEY"
      ]
    }
  ]
}

Set credentials in the gateway process environment:

export OPENAI_API_KEY=...
export BROWSERBASE_API_KEY=...
export BROWSERBASE_PROJECT_ID=...
export GEMINI_API_KEY=...
mcp-agentify --config /absolute/path/to/mcp-agentify.json

inheritEnv is an explicit allowlist. Backend processes receive a minimal default environment plus only listed variables and configured env values. A configured value such as "TOKEN": "${TOKEN}" expands from the gateway environment.

MCP client

Configure the gateway as a stdio MCP server:

{
  "mcpServers": {
    "agentify": {
      "command": "npx",
      "args": [
        "-y",
        "@steipete/mcp-agentify",
        "--config",
        "/absolute/path/to/mcp-agentify.json"
      ],
      "env": {
        "OPENAI_API_KEY": "..."
      }
    }
  }
}

The gateway exposes one MCP tool:

• orchestrate_task: selects and calls exactly one tool discovered from the configured backends.

Multi-step workflows require multiple orchestrate_task calls. For example, a Browserbase workflow can call start, then navigate, then extract.

CLI

mcp-agentify --config <path> [--frontend-port <port>|--no-ui] [--model <model>]

Environment overrides:

Variable	Purpose
OPENAI_API_KEY	Required OpenAI credential
OPENAI_BASE_URL	Optional OpenAI-compatible base URL
OPENAI_MODEL	Override openaiModel
MCP_AGENTIFY_CONFIG	Default configuration path
FRONTEND_PORT	UI port, or disabled
LOG_LEVEL	Pino log level
AGENTS	Comma-separated openai/<model> UI agents

Local UI

Set frontendPort or pass --frontend-port. The dashboard binds to 127.0.0.1 and shows backend status, redacted logs, MCP traces, configuration, and optional direct OpenAI chat.

Security

• Restrict filesystem backends to the minimum required directories.
• Keep credentials in environment variables; do not put them in JSON or command arguments.
• Only variables listed in inheritEnv are forwarded to a backend.
• Configuration, logs, traces, errors, and backend command arguments are redacted before display.
• Dashboard HTTP requests require a localhost Host; browser origins and WebSockets must be same-origin.
• The dashboard is local-only and has no authentication. Do not proxy or expose it.

Development

npm ci
npm run lint
npm test
npm pack --dry-run

npm test rebuilds the server and packaged UI before running unit and integration tests.

See API, examples, and release notes.