openai-agents

This MCP server exposes OpenAI agents as MCP endpoints so clients can orchestrate and delegate tasks to specialized tools. It provides three main capabilities: Web Search Agent for real-time web information, File Search Agent for searching and analyzing content in OpenAI’s vector store, and Computer Action Agent for performing safe, simulated computer actions. In addition, an Orchestrator Agent can coordinate between these specialized agents to decide which tool to use for a given task. Clients such as the Claude Desktop app can discover and invoke these agents through the MCP protocol, enabling composable workflows across agents.

To use the server, run it with an appropriate Transport (for example SSE) and provide your OpenAI API key. Once running, you can send MCP requests to the server to invoke specific agents or the orchestrator, supply any necessary inputs (like vector store IDs for file search or context for web search), and receive structured responses. The server is designed to be extensible, so you can add more specialized agents or adjust the orchestrator’s behavior to fit your workflows.

Prerequisites:

• Python 3.11 or higher
• uv package manager (recommended; install via pipx or system package manager)
• OpenAI API key

Option A: Install via Smithery (for Claude Desktop users)

1. Ensure Node.js and npm are installed.
2. Run:

npx -y @smithery/cli install @lroolle/openai-agents-mcp-server --client claude

3. Follow Smithery prompts to set up the client linkage.

Option B: Local development setup (manual)

1. Create and activate a Python virtual environment:

uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

2. Install dependencies and the MCP tooling (via uv):

uv sync --dev

3. Run the server locally with a chosen transport (e.g., stdio or sse):

export OPENAI_API_KEY=your-api-key
export MCP_TRANSPORT=sse
uv run mcp dev src/agents_mcp_server/server.py

4. Point your MCP client to the server URI (e.g., SSE endpoint) and begin sending MCP requests.

Environment variables: OPENAI_API_KEY is required. MCP_TRANSPORT can be stdio (default) or sse for server-sent events. The File Search Agent requires vector_store_ids corresponding to your OpenAI vector stores. The Computer Action Agent uses a simulated AsyncComputer implementation by default; to enable real interactions, implement and plug in a concrete AsyncComputer. If you plan to deploy, consider configuring additional environment variables for logging, rate limiting, and security. The server can be extended by adding more specialized agents or enhancing the orchestrator for complex workflows.