BytePioneer-AI/codex-host

CodexHost – Run Multiple AI Agent Harnesses Inside Codex Desktop

What it is

• A desktop‑side extension that lets you run many AI‑coding assistants (Pi, Claude Code, Grok, etc.) inside the official Codex Desktop IDE.
• Each assistant, called a Harness, appears as a separate session but shares a single project sidebar, diff view, and chat UI.

Key capabilities

Feature	What you get
Cross‑agent collaboration	Type # in the chat to hand a task to another Harness (e.g., #pi or #claude-code). The target runs in its own native session while you stay in the same workspace.
Unified diff & fork UI	Every response can show a diff of code changes, and you can fork a branch or a new worktree directly from any message.
Tool‑call visibility	Expand any edit or command to see the underlying tool call, reasoning steps, and token usage.
Remote harnesses	Run a Harness on a remote VPS via SSH (or experimental Windows Remote Control) while the UI stays on your local machine.
Live usage stats	Token count, cache‑hit rate, and cost estimate are displayed in‑app.
One‑click account import	Pull your locally signed‑in Codex/Grok credentials into other Harnesses like Pi, keeping quota info in sync.
Mermaid diagram rendering	Diagrams are rendered directly in the desktop UI, unlike the plain‑text TUI of some agents.

How to get started

1. Install via npm (npm install -g @codexhost/cli) or download the platform‑specific installer from the Releases page.
2. Run codexhost (or codexhost update to keep it current). The CLI launches Codex Desktop with the host layer attached.
3. Inside the IDE, open the bottom‑right Agent selector and choose any supported Harness. Use # to switch or delegate tasks.

Supported Harnesses (as of the README)

• Pi (and the Oh My Pi plugin)
• Claude Code
• OpenCode
• Grok
• DeepSeek Harness
• Antigravity CLI
• CodeBuddy, WorkBuddy, Cursor, Hermes, Qoder, Kimi Code, and others. All of them get core capabilities like streaming responses, tool status, edit diffs, slash commands, and cross‑agent task hand‑off.

Technical overview

• The host injects a CLI shim that sits between Codex Desktop’s native server and the Harness plugins.
• Each Harness is wrapped by a small adapter (Pi uses RPC, Claude Code uses the Agent SDK, others fall back to the generic ACP protocol).
• The UI is not rebuilt; CodexHost extends the existing Electron‑based desktop via CDP/Electron Inspector, so you get the official look and feel.
• Remote execution works by installing the same codexhost binary on a remote machine, starting a “remote” service, and connecting over SSH.

Who might use this

• Developers who already rely on Codex Desktop and want to experiment with multiple LLM coding assistants without juggling separate apps.
• Teams that need a unified view of code changes, token usage, and cost across different providers.
• Users who want to offload heavy compute to a remote server while keeping the UI local.

Community & contribution

• A WeChat group is provided for help and discussion.
• Contributions are welcomed; the repo includes a guide for adding new Harness adapters and a skill (codexhost-add-harness) to help developers scaffold plugins.

Bottom line CodexHost is a genuine, feature‑rich integration layer that brings many AI coding agents together inside the Codex Desktop environment, offering a seamless, collaborative workflow for developers working with LLM‑based code assistants.