A board that shows which AI agent is waiting on you.

English · 繁體中文 · 简体中文

Claude Code, Codex, Antigravity and other AI agents report to one board on your computer. When one stops to ask you something, its question goes to the top within seconds, so you stop hunting through windows. One Markdown file, no account.

Quick start

With Node.js 22+, in any terminal:

npx @wenyuchiou/agent-board@next setup

Then open a new session of your agent and run:

npx @wenyuchiou/agent-board@next

• setup shows what it will change and asks once before it writes.
• In that new session, type /agent-board ($agent-board in Codex; in Hermes, ask it for the Agent Board skill), or run the second command (Cursor has no such command).
• Per-agent steps, language, options and other ways to install: setup guide.

What you get

• Questions first. A question an agent is waiting on goes to the top, with a button for each option it can read, or you write your own answer.
• Every agent stays where it is. Agents keep running in their own terminals; the board reads what their hooks report.
• Local and plain. One Markdown file you can read and edit, a page on 127.0.0.1 only, and no model calls by the board.
• One command to install, one to remove. setup shows what it will change before it writes; remove takes out what setup wrote and leaves your own settings alone.

Supported agents

In a terminal (command line)

	Agent	Open the board	Your reply reaches it
	Claude Code	/agent-board	Sent when it is free
	Codex	$agent-board	Held for the session when its card offers it; when it is handed over
	Antigravity	/agent-board	Held for the session when its card offers it; when it is handed over
	Cursor	The second Quick start command	You paste it
	OpenCode	/agent-board	You paste it
	Grok Build	/agent-board	You paste it
	Hermes Agent	Ask it for the Agent Board skill	You paste it

In a desktop app or IDE

	Agent	Open the board	Your reply reaches it
	Claude Code	/agent-board (board page)	Not yet
	Codex	Board page	Not yet
	Antigravity	Board page	Held for the session when its card offers it. Desktop app: checked in a live run; IDE: no run of its own, and it is treated like the app when its hook reports it as the app's surface; when it is handed over
	Cursor	Board page	Not tested yet
	OpenCode	Board page	Not yet

• Board page: run npx @wenyuchiou/agent-board@next --html in a terminal, the app's own included.
• Not yet or Not tested yet: copy your answer from the board and paste it into the app.
• dot and other web chats report by hand: how.

Answer from the board

An answer to an agent's question and a message you send it travel the same way, by the button you press:

• Send hands it to a connected session; a confirmation from that session means it was received. With no confirmation it may still have arrived, so look at the agent before you use Copy answer, or it may get it twice. Send needs messaging on.
• Queue for the agent's next turn holds it for that session's turns, does not wake an idle one, and can be rejected if its 12-hour mailbox expires or, for an agent whose hook reports the end of a session, if the session ends first.
• Copy answer puts it on your clipboard to paste into that agent.
• Answered in the agent's own window? Press Already answered, hide to clear the card.

Open it anywhere

Where	How
In an agent	/agent-board ($agent-board in Codex)
A terminal	npx @wenyuchiou/agent-board: the terminal board in an interactive terminal over SSH or on Linux with no display, unless an agent's setting such as HERMES_HOME or OPENCODE_CONFIG_DIR is set in that shell; plain text in CI; the page everywhere else. --watch gives the terminal board in any interactive terminal
A phone on the same Wi-Fi	npx @wenyuchiou/agent-board --lan and scan the QR code. The phone can look, tick, archive and close the board, but cannot send an answer (it shows the answer to copy). Use it on a network you trust: the link is not encrypted and anyone who has it can open the page
An agent reading the board	npx @wenyuchiou/agent-board --json (fields)

How it works

Each agent's hooks write what its session is doing into its own block of one local progress.md (a fresh install uses ~/.agent-board/progress.md; the format). The page listens on 127.0.0.1 only and needs no account. The board does not call a model; what the optional Claude Code plugin can send out is in the guide.

Update and remove

Run npx @wenyuchiou/agent-board@next setup again to see what an update would change, then setup --apply; files you edited are left alone. To remove what setup wrote:

npx @wenyuchiou/agent-board@next remove
npx @wenyuchiou/agent-board@next remove --apply

Your history stays beside progress.md; Update and uninstall lists what is kept, and a few small cache files can be left behind. For the shorter agent-board command, run npm install -g @wenyuchiou/agent-board@next once.

Troubleshooting

You see	Try
An agent is not on the board	Start a new session of it; one that was already open is not connected. For Codex, preview npx @wenyuchiou/agent-board --codex-trust --codex-bin "<path-to-codex>" --json with Codex's full path, review it, then add --apply only if you approve. Codex CLI also offers /hooks.
Nothing changed after installing	Run npx @wenyuchiou/agent-board@next setup again without --apply (from a checkout: node scripts/setup-agent.mjs). Read the conflicts and Next steps; edited files are left alone and listed there.
If the board looks old	A session that was already open may still run the version it loaded: start a new session of that agent. To update, see the guide.
An agent shows up twice	It may be connected both by the Claude plugin and by setup: what to remove.

To check Claude Code's setup, run claude plugin list and /progress connect status. More answers: the guide's FAQ and Troubleshooting. Found a bug? Open the bug form and follow the safe-report steps: no session ids, home-folder paths or tokens.

Documentation

Read	For
Guide	Daily use, keys, the page, settings, update and uninstall, FAQ
Setup	What the installer writes for each agent
Codex · Messaging · File format · Remote reporters	Reference
AGENTS.md · Contributing · Security	Working on the code; how a non-Claude agent reports by hand

License

MIT. Agent logos are from Simple Icons (CC0) and Lobe Icons (MIT) and belong to their owners; an agent without a logo is shown by its initial.

_{Latest stable version: 0.24.2 · 0.24.3 is a pre-release (npx @wenyuchiou/agent-board@next setup) · Changelog.}