I build Elozi, a customer-service workspace. I wanted AI assistants to help with support without giving them free rein, so I published elozi-mcp (MIT, v0.1.1).
It is a small stdio bridge that connects desktop agents (Claude Desktop, Claude Code, Cursor, Codex or any MCP stdio client) to Elozi's hosted endpoint. It stores no data locally, adds no telemetry and does not retry tool calls automatically.
Permissions come from the token. In the Elozi panel you create an agent of type "Agent via MCP" and a service identity with channels, departments, capabilities, validity and daily quotas. The assistant only sees tools your token allows. A good start is conversation.read plus reply.submit: it reads and drafts, a person approves. Add reply.send or conversation.manage only when you want it to act alone.
Requirements: Node.js 22+ and an ELOZI_API_TOKEN (starts with elozi_svc_, shown once).
Install / run:
npx -y elozi-mcp Cursor example (~/.cursor/mcp.json):
{ "mcpServers": { "elozi": { "command": "npx", "args": ["-y", "elozi-mcp"], "env": { "ELOZI_API_TOKEN": "elozi_svc_YOUR_TOKEN" } } } } Claude Code users can also connect over HTTP directly:
claude mcp add --transport http elozi https://elozi.je4ndev.com/api/agent/mcp --header "Authorization: Bearer ${ELOZI_API_TOKEN}" Keep the token in private config only, never in chats or URLs.
npm: https://www.npmjs.com/package/elozi-mcp
Repo: https://github.com/JE4NVRG/elozi
