MCP 服务器

Omega API MCP

It gives AI agents one tool to access the Omega-API judge, which evaluates conversational turns and returns a decision and gate action (release, clarify, suppress, or none) to guide the agent's response.

README

@dailui/omega-api-mcp

An MCP server that gives an AI agent one tool to reach the Omega-API judge.

Give agents access to the judge without moving the judge into the agent.

The agent calls the tool with a turn; the judge — which stays on the server — returns a decision and a gate action. Zero judge logic runs in the agent. Apache-2.0. Node 18+.

The tool

omega_judge_turn — judge one agent turn: does the conversational process move? Returns a decision and a gate action. It measures process shape, not whether the goal was reached.

Input:

field	type	notes
`input`	string (optional)	the user / incoming message
`output`	string (optional)	the agent's drafted reply (at least one of these)
`session_id`	string (optional)	one id per conversation
`privacy_mode`	`FULL` \| `REDACTED`	optional

Result:

{
  "turn_id": "t_…",
  "decision": "PROCEED | CLARIFY | STOP",
  "reason_family": "…",
  "side": "…",
  "gate": { "action": "release | clarify | suppress | none", "release_original": true }
}

Act on the gate before showing the reply: release → show · clarify → regenerate then call again · suppress → stop or escalate · none → observe only, show.

Before you start: register → key → judge → meter

Register a project and get an API key (oapi_live_…) at https://omega.dailui.com/signup.php. Your first 100 judgments are free; after that, billing applies.

Add it to your MCP host

Most hosts (Claude Desktop, IDE agents) launch MCP servers from a config file. Add:

{
  "mcpServers": {
    "omega-api": {
      "command": "npx",
      "args": ["-y", "@dailui/omega-api-mcp"],
      "env": { "OMEGA_API_KEY": "oapi_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" }
    }
  }
}

The host starts the server over stdio and the omega_judge_turn tool becomes available to the agent.

What it does not claim

It measures whether the process moved. It does not promise the agent achieved the user's goal — the two are independent. It returns no scores or internal measures, only the decision and the gate.