Run AI Agents from Claude Code, Cursor, and Windsurf with OE MCP

OE MCP v1.6.6 ships a new tool — run_agent — that lets Claude Code, Cursor, Windsurf, Codex, and any MCP-compatible AI app trigger OE Runtime YAML agents directly. No terminal. No separate process. Just ask.

What You Need

• OE MCP installed — run npx -y @openenthrium/oe-mcp or follow the setup guide
• Claude Code, Cursor, or Windsurf with MCP support enabled
• At least one OE Runtime agent YAML in your project folder
• An API key from OpenAI, Anthropic, or any supported LLM provider

Why This Matters

OE Runtime lets you define multi-step AI agents in plain YAML — connecting to databases, SSH servers, APIs, Slack, Telegram, and more. Until now, running those agents required a terminal command.

With run_agent, your AI coding assistant becomes the trigger. You describe what you want in plain language, and Claude (or Cursor, or Codex) figures out which agent to run and calls it directly through MCP. The full output comes back into the conversation.

The Tool

Under the hood, OE MCP runs:

npx -y @openenthrium/oe-runtime@latest ./skill-folder-name [--param key=value ...]

Config Auto-Detection

Each agent directory can have its own oe-config.json with the correct LLM credentials and connector names for that specific agent. OE MCP auto-detects it:

• Looks for oe-config.json inside the skill folder (alongside SKILL.md and agent.yaml)
• Uses it if found — correct LLM + connector config for that agent
• Falls back to the main oe-mcp.json config otherwise

This means each agent can target a different database, LLM provider, or set of connectors — without any conflicts.

Real Example — Database Analyst Agent

Here's a DB analyst agent that queries a PostgreSQL database and sends a summary to Telegram. Create a db-summary/ folder with two files:

# db-summary/SKILL.md
---
name: db-summary
description: Queries a database and sends a summary to Telegram
license: Apache-2.0
metadata:
  author: Open Enthrium
  version: "1.0"
---

You are a database analyst agent. Query the database, summarize findings,
and send the summary to Telegram.

## Step 1: Query Database
Query the workspaces table and summarize what you find.
GET all rows and note the count, names, and any key fields.

## Step 2: Send to Telegram
Get the latest chat_id from GET /getUpdates.
Send a clean summary via POST /sendMessage.

# db-summary/agent.yaml
name: DB Summary Agent
description: Queries a database and sends a summary to Telegram
connectors:
  - connection_name: My Database
    connection_type: postgresql
  - connection_name: My Telegram Bot
    connection_type: telegram
skills:
  - path: ./
    trigger_type: auto

To run it from Claude Code, just ask:

You: "Run my database analyst agent at E:/oe-runtime/sql-databases/db-summary/"

Claude calls run_agent → OE Runtime executes → result returned:

🔧 My Database
   ↳ [{ "id": 1, "slug": "ingest-a8c79e", "name": "Assistants", ... }]

🔧 My Telegram Bot (getUpdates)
   ↳ { "ok": true, "result": [{ "message": { "chat": { "id": 5945555985 } } }] }

🔧 My Telegram Bot (sendMessage)
   ↳ { "ok": true, "result": { "message_id": 14 } }

✅ Done — workspace summary sent to Telegram (message_id: 14)

Passing Params

Agents can declare required params in their YAML. Pass them via the params object:

# Claude Code prompt:
"Run my report agent for the month of July"

# Claude calls:
run_agent(
  path: "/agents/report/",
  params: { "month": "July", "format": "pdf" }
)

# OE MCP executes:
npx -y @openenthrium/oe-runtime@latest /agents/report/ \
  --param month="July" \
  --param format="pdf"

What You Can Build

Setup in 3 Steps

1. Install OE MCP v1.6.6+

# Claude Code / Cursor / Windsurf — .mcp.json (macOS/Linux)
{
  "mcpServers": {
    "oe-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@openenthrium/oe-mcp", "--stdio", "/path/to/oe-mcp.json"]
    }
  }
}

2. Create your agent directory

my-agent/
  agent.yaml       # the agent definition
  oe-config.json   # LLM + connector credentials

3. Ask Claude to run it

"Run my agent at /home/user/my-agent/agent.yaml"
"Run the report agent for last month"
"Execute the SSH diagnostics agent on prod-server"

Combining with Other OE MCP Tools

Because run_agent is just another MCP tool alongside your connectors and memory, Claude can orchestrate all of them in a single conversation:

• Query a database directly with a connector tool, then run an agent to send results to Slack
• Store a value with memory_set, then run an agent that uses it as a param
• Check the action log with log_list to see what connectors the agent called

The AI app becomes a full orchestration layer — not just a code editor, but a runtime controller for your entire agent library.