devilcoolyue / agentbox

# agentbox
**Your AI coding workspace. On your server. In your browser.**
Run Claude Code and Codex CLI on a Linux server you control. Go from a task to reviewed code with chat, terminal, files, and Git in one place.
**English** | 简体中文
Watch the tour · Highlights · Screenshots · Quick start · Architecture · Documentation
## See it work
 ▶ Full tour in 1920 px (MP4, 70 s) — adds file previews, usage records and switching between the six interface styles · Screenshots 
 Recorded from the real web console with synthetic projects, conversations, command output and usage — no model was called. It demonstrates the interface, not model quality or speed. 
## Why agentbox?
Set up your coding environment once on your server, then pick up your project from a laptop, desktop, or phone. Your local machine only needs a browser; the agents and project tools run in Docker on the server.
| What you need | What agentbox provides |
| --- | --- |
| **Pick up where you left off** | Persistent project files, home configuration, and chat history across browser sessions |
| **Work with familiar agents** | Claude Code and Codex CLI, with a streaming Web chat and their original interactive terminal interfaces |
| **Finish the task in one place** | Upload or clone → chat → run and preview → review diffs → commit, push, or download |
| **Reuse your setup** | User-level shared files, layered home templates, Claude skills, and MCP configuration with workspace overrides |
| **Give a trusted team access** | Central account pools, per-user account scopes, per-request usage records, price snapshots, and optional credits |
| **Connect to your own services** | Optional abox-link tunnel to allowlisted repositories, databases, and other private-network services |
You bring the server and your own Claude / Codex subscription, API key, or compatible relay account. Agentbox provides the workspace and management layer; model usage follows your provider's terms and pricing.
## Highlights
### A chat that keeps up with the CLI
- **Watch the work, not a spinner.** Replies stream with thinking summaries and every tool call; Codex commands are named the way its CLI shows them (Read, Search, Shell, Edit) with exit status and duration, and their output and diffs are a click away. The agent's task list sits above the message box ("Tasks 2/3 · Running the tests"), a timer counts the turn, and a ring shows how full the context window is, like the CLI status line.
- **`@` files and `/` commands.** Type `@` to search workspace files by name and insert the path. Type `/` to list the workspace's real Claude Code commands. `/context` is drawn as the CLI's colored grid and, on subscription accounts, `/usage` as limit bars; both are answered locally without a model call.
- **Stay in the loop.** When Claude asks a question mid-task, answer it in a card and the same turn continues. Keep typing while it works: messages queue for the next turn, or steer the running one. After a turn, the predicted next prompt