MCP server overview
Connect an AI assistant to your agent — query the Knowledge Base, read analytics, add files, and reach workers, all with a human approving each send.
The OnKey MCP server exposes your agent's capabilities as Model Context Protocol tools, so an AI assistant (Claude, ChatGPT, or any MCP-compatible client) can work with your agent on your behalf — using the same data and pipelines your Virtual Operations Agent uses when you email it.
It's the sibling of the REST API: same connection, same Agent API Token. Use the MCP server when an AI assistant is doing the work; use the REST API when your own code is.
Endpoint
https://mcp.onkey.chat/functions/v1/mcp-serverIt speaks MCP Streamable-HTTP JSON-RPC over HTTPS. Any MCP client connects normally.
What you can do
- Query the Knowledge Base and get grounded answers with citations.
- Read analytics — usage stats and recent worker questions.
- Manage content — add files to the Knowledge Base, preview and apply Prompt Card edits.
- Reach workers — broadcast Alerts, send targeted Messages, schedule Reminders, and reply to worker questions.
Which of these your token can do depends on its scope (read or full).
The key difference from the REST API
Every worker-facing action over MCP goes through a stage → confirm gate: the assistant stages the action, a human reviews an exact preview, and only then does it send. This exists because an AI assistant can misread an instruction. The REST API has no such gate — there, a developer wrote the exact request. See Safety: stage → confirm.
Next steps
- Connect your client — Claude Desktop / Claude Code config.
- Token scopes — what
readvsfullunlock. - Tool catalog — every tool.
- Safety: stage → confirm — how sends are gated.
Updated about 1 hour ago