OpenChatX turns regular ChatGPT into a local agent that uses your normal Chat allowance instead of Codex minutes

A developer on a ChatGPT Plus subscription who wants a coding agent to touch their machine usually has to move to Codex or the ChatGPT Work agent, each with a separate quota that runs out on its own schedule. Once that quota is empty, the ordinary Chat window is still there with unused messages, and none of it can reach the filesystem.

OpenChatX, released this week by a developer whose GitHub handle is XiaoPuOuO, connects the plain Chat window to a local agent runtime and spends only that Chat allowance. The MIT-licensed repository was created on 23 September and has 131 stars four days later. The desktop app is signed and notarised on macOS; the Windows build is unsigned and only offered from the GitHub Releases page.

How the connection works

The README says OpenChatX does not reverse-engineer ChatGPT, reuse browser cookies or intercept its traffic. Instead it uses OpenAI's documented Secure MCP Tunnel and Developer Mode MCP apps. A local MCP server runs on the user's machine, OpenAI's official tunnel-client opens an outbound HTTPS connection to a Secure MCP Tunnel, and ChatGPT connects to that tunnel as a custom MCP app. Tool calls go through the same OpenAI channel to the local runtime and back.

The setup steps are on the OpenAI Platform side (create a tunnel), the ChatGPT side (enable Developer Mode, create the MCP app) and the OpenChatX desktop app (paste the tunnel ID and a Runtime API key). ChatGPT stays the planner. OpenChatX is the tool layer.

What ChatGPT gets on the other side

The runtime exposes coding-grade file and shell tools, MCP server aggregation, portable Skills as SKILL.md workflows, .mdc Rules in Cursor's format, an editable AGENTS.md template, folder-bound Projects, and a Subagent layer that lets the main agent delegate to explicitly configured model profiles. External MCP schemas stay lazy so ChatGPT's tool context does not blow up.

The feature the README calls a differentiator is cross-session Summaries. When a chat approaches its practical context limit, the agent calls summarize, which stores a compacted older-context summary and roughly the latest 8,000 tokens of recent conversation locally, and returns a UUID. Opening a fresh ChatGPT conversation and handing over the UUID restores the two pieces, then the handoff is deleted.

The caveats worth stating

OpenChatX runs with the user's normal permissions on their own machine. The README's caution block says a trusted ChatGPT caller can run commands, edit files, use connected MCP servers and control supported applications. The security notes say the local MCP endpoint is meant to stay on loopback, the Secure MCP Tunnel is the supported remote path, and tool and audit logs may contain sensitive input. Only trusted MCP servers should be connected.

Two smaller cautions: the Windows binary is not Authenticode-signed, so Windows will warn on install, and a manual source install needs Node.js 22.18 or newer, ripgrep and the OpenAI tunnel-client. Everything else, including the tunnel-client, is bundled in the desktop app.

The interesting piece here is the routing. ChatGPT is the model most people already pay for, and until now it could not touch a laptop without moving the work to Codex, Work or the Atlas agent, each with a different quota. Pointing the messages a user already has at a documented MCP tunnel turns the same subscription into a coding-agent seat, without the reverse-engineering risk the earlier ChatGPT-agent projects on GitHub carried.

Source

Primary source: XiaoPuOuO/openchatx-mcp on GitHub. Referenced OpenAI documentation: Secure MCP Tunnels and Developer Mode and MCP apps in ChatGPT.