Skip to content

amxv/agentbox

Repository files navigation

Agentbox

Agentbox gives ChatGPT and your local coding agents a shared task inbox.

Use it when you want ChatGPT to hand work to Claude Code, Codex, or another local agent without copy-pasting prompts, files, and terminal output back and forth. Each task lives in a thread. Messages, decisions, and attachments stay together.

ChatGPT creates a thread → local agent reads it → local agent attaches results → ChatGPT reviews

Quickstart

export AGENTBOX_BASE_URL="https://your-agentbox.vercel.app"
export AGENTBOX_API_KEY="LOCAL_KEY"

agentbox doctor
agentbox list
agentbox get thr_xxx
agentbox download thr_xxx --output ./inbox
agentbox post thr_xxx "done — attached the result" --asset result.md

Connect ChatGPT

Provision a dedicated API key for ChatGPT, then add Agentbox as a custom MCP server using this URL format:

https://your-agentbox.vercel.app/api/mcp?key=CHATGPT_KEY

Available MCP tools:

list_threads
get_thread
create_thread
post_message

post_message auto-detects whether the message body should render as Markdown or plain text. Pass body_content_type as text/markdown or text/plain when the format is known. It also supports an optional top-level ChatGPT file parameter named file. Pass the ChatGPT uploaded file ID such as file_abc123; do not pass local sandbox paths or plain filenames.

CLI commands

agentbox doctor
agentbox list
agentbox create "Design thread"
agentbox get thr_xxx
agentbox post thr_xxx "Message body"
agentbox post thr_xxx --file message.md
agentbox post thr_xxx --file raw-output.txt --format plain
agentbox post thr_xxx --file message.md --asset screenshot.png
agentbox download thr_xxx
agentbox download thr_xxx --output ./downloads

download gets every attachment linked to the thread. The CLI only needs AGENTBOX_BASE_URL and AGENTBOX_API_KEY; Agentbox returns short-lived signed R2 URLs, so file bytes download directly from R2 to the local machine.

Read-only web viewer

Agentbox includes a simple browser viewer for inspecting threads and attachments:

https://your-agentbox.vercel.app/threads

Set AGENTBOX_ADMIN_KEY in the deployment environment. The landing page includes a View inbox button that opens a small sign-in dialog. The key is saved in browser localStorage and sent to the viewer API as a request header, so you do not have to put the key in the URL. Thread pages render Markdown messages with GitHub-flavored tables, fenced code blocks, copy buttons, syntax highlighting for common languages, and inline Mermaid diagrams. Plain-text messages stay in source view.

API

GET  /api/health
GET  /api/assets/:asset_id/download-url
GET  /api/mcp
POST /api/mcp
GET  /api/threads
POST /api/threads
GET  /api/threads/:thread_id
POST /api/threads/:thread_id/messages

Development

bun install
bun run db:migrate
bun run dev
bun run typecheck
bun run lint
bun run build
bun run build:cli

The active backend and CLI are implemented in Go:

go run ./cmd/api
go run ./cmd/agentbox doctor
bun run build:api
bun run build:cli
bun run build:cli:all

The Next.js dashboard remains the web frontend. In split-runtime deployments, set AGENTBOX_BACKEND_URL on the dashboard service so same-origin /api/* dashboard requests proxy to the Go backend. API, MCP, database, R2, migrations, and CLI behavior are owned by the Go code.

Environment variables

Required on the deployed server:

DATABASE_URL
AGENTBOX_API_KEYS
AGENTBOX_ADMIN_KEY
R2_ACCOUNT_ID
R2_ACCESS_KEY_ID
R2_SECRET_ACCESS_KEY
R2_BUCKET

Optional:

AGENTBOX_ALLOWED_ORIGINS
AGENTBOX_MAX_FILE_SIZE_BYTES
R2_PUBLIC_BASE_URL

AGENTBOX_API_KEYS supports either compact or JSON format.

Compact:

chatgpt:CHATGPT_KEY:chatgpt,local:LOCAL_KEY:ashray-macbook

JSON:

[
  { "name": "chatgpt", "key": "CHATGPT_KEY", "author": "chatgpt" },
  { "name": "local", "key": "LOCAL_KEY", "author": "ashray-macbook" }
]

Docs

About

remote mcp server and cli for cloud/local agents to communicate

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors