The Codex Cloud SWE Agent still doesn’t support Model Context Protocol (MCP), which means its cut off from the growing ecosystem of MCP tools. I built a tiny CLI wrapper, remote-mcp-cli, that fixes this with one command:
# Discover tools on any remote MCP server
uvx remote-mcp-cli https://mcp.deepwiki.com/mcp list
# Invoke a tool with JSON arguments
uvx remote-mcp-cli https://mcp.deepwiki.com/mcp call read_wiki_structure '{"repoName":"openai/openai-python"}'
How it works
- Terminal bridge – Codex (or any shell-capable agent) calls remote-mcp-cli, which works with modern Streamable HTTP MCP.
- Zero setup – No local server, no config files. Just add the MCP endpoint URL to call tools.
- Auth & headers – Supports bearer tokens and arbitrary headers.
- Tool discovery – list shows each tool’s schema so agents can auto-generate the right JSON payloads.
- AGENTS.md – Explain how to use the tools and provide some examples of commands that call your custom MCP tools in your AGENTS.md file. Codex reads that and treats the tool set just like any first-class MCP-compatible agent (like Cursor or Claude Desktop).
https://github.com/zueai/remote-mcp-cli
I’d love feedback, bug reports, and PRs. If you wire this into your Codex workflows (or any other agent), ping me—keen to hear what you build!