Codex MCP setup for rConfig: connect OpenAI Codex CLI to your network configs

Codex MCP setup for rConfig: connect OpenAI Codex CLI to your network configs

OpenAI’s Codex CLI connects to rConfig’s MCP server over Streamable HTTP, reading the bearer token from an environment variable. For teams on OpenAI models it is the way to use rConfig with an agent, because ChatGPT can’t connect.

Best for: OpenAI-based teams, terminal work, scheduled checks on a server or CI runner.

Connect

Add rConfig to ~/.codex/config.toml, then set the variable in the environment Codex runs in:

[mcp_servers.rconfig]
url = "https://<your-rconfig-server>/mcp"
bearer_token_env_var = "RCONFIG_MCP_TOKEN"
default_tools_approval_mode = "writes"

export RCONFIG_MCP_TOKEN="..."

codex mcp add rconfig --url https://<your-rconfig-server>/mcp creates the table too, but has no flag for the token, so add bearer_token_env_var by hand. The MCP server page is the reference for the connection settings.

Check it: run codex mcp list, or /mcp inside a session, and confirm rconfig and its eight tools.

Approvals

default_tools_approval_mode = "writes" makes Codex prompt only for tools that aren’t marked read-only. rConfig marks all seven of its tools read-only (the MCP readOnlyHint), so they run without stopping. Two related settings for the same table:

• enabled_tools and disabled_tools limit which rConfig tools Codex loads, if you want a narrower agent.
• tool_timeout_sec (default 60) is ample: rConfig’s config search stops itself after about five seconds.

Things to ask

• Estate search: “Using rConfig, find every device with ip http server in its latest config. Read each matched line, because no ip http server matches too, and list only the devices where it’s enabled.”
• Script against reality: “Write a Python script that exports every device in the Branch group with its last backup time. Check the field names against rConfig’s get_devices output first.”
• Explain history: “Show the last five config versions of edge-fw-02 in rConfig and summarise how its NAT rules evolved.”

Run it unattended

codex exec runs a prompt without the interactive interface, so a cron job or CI pipeline can produce a report. Use a token for a dedicated rConfig user scoped to the devices the job covers, and keep the approval mode at writes so read-only calls don’t wait for a person.

Data handling

Tool results, including configs with secrets obfuscated, go to OpenAI’s models as part of the conversation. Check that your OpenAI agreement suits your configuration data before connecting production devices.

Related reference

• Connect an AI agent over MCP
• Which AI agents can connect to rConfig?
• ChatGPT and Claude on the web
• Find devices exposed to a security advisory