Agentrove: A Unified Self-Hosted AI Programming Workbench for Claude and Codex

Agentrove is a self-hosted AI programming workbench that supports both Claude Code and OpenAI Codex dual engines, designed to solve the fragmentation problem of AI programming tools. It integrates the two engines via the ACP protocol, provides isolated sandbox environments for running agents, and integrates a complete development toolchain including chat, editor, terminal, Git, etc., ensuring data sovereignty while improving development efficiency.

In the current AI programming field, Anthropic's Claude Code (excellent at code understanding and long-context reasoning) and OpenAI's Codex (outstanding in code generation and API usage) stand side by side as two leading options, but developers are often forced to choose one or switch between tools. Agentrove integrates the two engines into a unified interface via the ACP adapter and supports self-hosting, preventing sensitive code from flowing to third-party servers—making it especially suitable for enterprise users.

Agentrove adopts a layered architecture: the frontend uses React19/Vite to provide IDE-style panels (chat, edit, terminal, etc.); the backend uses FastAPI, supporting Web (PostgreSQL/Redis) and desktop (SQLite/in-memory cache) modes; the sandbox layer provides independent Docker or host environment isolation for each workspace; the agent layer interacts with the dual engines via the ACP protocol. Workspaces support multiple source types (empty directory, Git clone, local folder). The same workspace shares a sandbox, different workspaces are isolated, and the Git worktree mode supports parallel task processing.

Agentrove supports both Claude series (Sonnet/Opus/Haiku) and Codex series (GPT-5.4, etc.) models, allowing developers to choose based on tasks. It provides fine-grained control: permission modes (from read-only to full access), inference depth (balance between speed and quality). The built-in toolchain includes Monaco editor (same as VS Code), xterm.js terminal, diff comparison, full Git workflow integration (branch management, PR generation), and key management (secure injection into sandbox).

Deployment options: Web mode (one-click deployment via Docker Compose, suitable for team collaboration), desktop mode (macOS app built with Tauri, runs locally, data stored locally). Session management supports streaming responses (real-time viewing of AI thinking), resumable sessions (breakpoint resume), and sub-thread mechanism (branch processing of side issues, conclusions merged back to the main thread).

Agentrove integrates the advantages of dual engines, provides self-hosting and a complete toolchain, and adapts to both individual and team scenarios. However, note the following: the project is in an active development phase, so versions may have breaking changes; running dual engines requires sufficient API quotas and computing resources; currently, it mainly focuses on macOS desktop applications, and support for other platforms is pending expansion.