A console for developers running several agents across several projects at once. Stable, private, quick, and good to look at.
English · 简体中文
vibepanel is a working terminal for people who lean on agents hard. The front end and the back end are separate: your terminals live in a dedicated system service with a memory floor and a high priority, so an OOM or a crash somewhere in the application layer does not cost you a session. The interface is a web page, which means you can point it at a development box and reach it from whatever device is in your hand.
On the security side, every line of the source is public and every release is packaged by a public GitHub Action. What comes out is one Go binary with no dependencies that never phones home, which is about as much as I can do for safety, portability and speed at the same time. Upgrading or restarting the panel does not touch a single session or a single running agent. It terminates its own TLS, so you can put it on https and require a password or a passkey to get in.
On the UI side there are a lot of small deliberate things. The whole model is projects and sessions, so you can see at a glance what every session in every project is doing: finished, working, or waiting on you. Down the right there is a file manager and a notebook, and you can move files and images in and out by copy and paste, the way you would on your own machine.
Along the bottom I added a scratch terminal, because that is how I work: it is there to look at a file or run something while the agent is still going, and it saves you from interrupting the agent to ask. There is a separate interface built for phones, and you can turn on system notifications or point them at a channel of your own. Leaving the house stops being a reason to stop shipping.
And you do not even need the page open. Pair the panel with a Telegram bot, a
飞书 app or 微信 and a session that wants you sends a card there: which session,
what it is waiting for, and on Telegram and 飞书 an Allow and a Deny button
right under it. Reply 3: run the tests again and the words go into session 3;
shot 3 sends back a picture of the pane. Turn on the advanced mode and you can
just say "tell the docs one to add a changelog entry", and a headless Claude
Code or Codex works out which session you mean, then waits for your ok before
it types anything.
The panel knows when an agent is actually done, not only when it has gone
quiet: one click installs a hook into Claude Code, Codex, Kimi Code, zcode or
opencode, showing you what it will write and backing the file up first. And
upgrading is a button too: Settings → Updates downloads the release, checks it
against SHA256SUMS, runs the new binary once before trusting it, swaps it in
and restarts, with every session still running afterwards.
The other thing I am fond of is share links. Put the whole system's status on a
monitor, or put tokens spent against code produced on a big screen so whoever
needs convincing can see it. What the screen shows is a page you have an agent
write, in its own project, with a live preview beside the terminal: start from a
template, publish, make a link, open it on the screen. Click the preview and it
opens large over the whole window, on a phone, a laptop or a television, so you
see what the wall will see before you hang it. Pages live beside the panel's own
data unless you say otherwise, and travel between panels as a zip. Fine, I
know this is a niche feature. I still think it matters.
The other thing I use every day is sessions in the chat app on your phone: Telegram, 飞书 or personal 微信, private chats only. When an agent stops to ask permission for a command or asks you a question, a card arrives that starts with the session's number, and you answer in the same chat: "3 yes", "3: add a test", or the Allow and Deny buttons on Telegram and 飞书. It understands the way people actually type on a phone, so leaving your desk stops being a reason for an agent to wait.
I take the privacy side of this seriously. The three chat integrations are written in-house with no third-party SDK and ship inside the same binary; the 微信 sign-in QR code is drawn locally in your browser, and the font screenshots are rendered with is bundled too, so nothing is fetched at run time. Until you configure and switch on a channel, the chat feature connects to no outside service at all, and a test holds it to that. The first time you switch one on, the panel tells you plainly what will pass through whose servers and waits for you to agree. A stranger who messages the bot gets a pairing code and nothing else until you enter it on the panel, and every "allow" applies only to the request you actually saw, so an old card can never approve a new command.
I will say this plainly: this is not AI slop. It is a terminal I use hard every day and it is genuinely nice to use. I hope it saves you some time and that you enjoy it. It is early, so come and join in if you have ideas or complaints.
Share pages: an agent writes it beside a live preview you can open large, you publish it, and a link puts it on a screen. How.
A separate interface built for phones.
What every project and model has spent, and what every session is costing the machine.
Anyone who has several terminal agents open at once and needs to keep track of them, or who has a development machine they want working around the clock and reachable from any device.
It speaks 简体中文 and English and supports several ways of installing. The comparison between them, and unattended installs, are further down.
curl -fsSL https://raw.githubusercontent.com/jiangmuran/vibepanel/main/install.sh | shThis is a public GitHub mirror run by jiangmuran. To keep it from being abused, the first run prints a prompt in your terminal asking you to pass a check in a browser.
curl -fsSL https://github.muran.tech/https://raw.githubusercontent.com/jiangmuran/vibepanel/main/install.sh -o vibepanel-install.sh \
|| curl -sSL https://github.muran.tech/https://raw.githubusercontent.com/jiangmuran/vibepanel/main/install.sh
sh vibepanel-install.sh --mirrorvibepanel service status | start | stop | restart | logs | token | upgrade | uninstallThe rest is in docs/install.md: user unit against system unit, every flag, unattended installs, creating the first account from the command line, Docker, and building from source.
docs/features.md — including
sessions in a chat app through
Telegram, 飞书 or 微信, with replies going back into the session, and
plugins: themes, panels, services and processes,
each installed from a screen that says in words what it may do; a process
can be served on the panel's own port behind the panel's door, and declare
the models and runtimes it needs for the panel to fetch and verify.
vibepanel with a subcommand operates the running panel from a shell --
projects, sessions, notes, todos, tokens, the activity log, share pages,
plugins -- through the panel's own API, so a rename from docker exec is
the same rename as from the sidebar. docs/cli.md lists the
commands; vibepanel --help does too.
Every flag has a VIBEPANEL_<UPPER_SNAKE> environment equivalent, and the flag
wins. A VIBEPANEL_* variable nothing reads is reported at startup and by
doctor rather than ignored, so a renamed setting is loud instead of quietly
doing nothing. The full table is in docs/install.md.
The same binary is also the admin CLI: serve, project, session, hook,
service, account, doctor and version.
TOKEN=… # Settings → API tokens
curl -sX POST https://panel.example.com:18443/api/sessions \
-H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
-d '{"projectId":"…","title":"billing","command":["claude"]}'command is an argv; leave it out and you get a shell, which is what the
panel's own interface sends. Everything the front end can do is available over
the same API. docs/api.md is the whole surface, and it is checked
against the router in both directions: an endpoint missing from the page and a
paragraph describing one that no longer exists both fail the build.
The principles the rest follows from:
- The page is a view, not the state. Close it, open it in three places, reload it mid-command; the sessions never notice.
- Done means the process exited, not that a session went quiet.
- Colour is never the only carrier of meaning.
make check # vet, gofmt, eslint, Go tests, frontend units — the fast gate
make verify # everything, including the browser checks (~20 min)
make head-check # build and test a clean worktree at HEAD, not the working treemake check never starts a browser. Most of this project's bugs were found by
the ones that do:
make panes-check |
the side panel's pane layout: drag, drop, merge, restore |
make first-run-check |
the setup wizard and the first project |
make archive-check |
archiving a project: the sidebar, the list, the picker, a phone |
make render-check |
layout, states, arbitration, panels, mobile, clipboard, passkeys |
make stress-check |
wide characters, full-screen programs, scrollback, floods, dropped sockets |
make restart-check |
kill the backend; the sessions and the login must outlive it |
make scale-check |
two dozen sessions: snapshot size, sidebar reachability, poller |
make pages-check |
share pages: the sandbox from inside a page, the editing loop, every template |
make plugins-check |
plugins: arrive, the install screen in both languages, a grant withheld, settings, a theme, a new plugin from a template, layout at three widths |
make tls-check |
its own TLS: wss, the Secure cookie, swapping a certificate |
make install-check |
both installers down every branch, in both languages |
make release-check |
build the archives and run one from a throwaway HOME |
The tmux wrapper is tested against a real tmux on a throwaway socket rather than
a mock; TEST_TMUX_BIN=/path/to/tmux go test ./... points it at another build.
web/scripts/shots.mjs takes screenshots by booting the real binary and
photographing it.
AGENTS.md has the conventions and the red lines.
PolyForm Noncommercial 1.0.0, with attribution terms. Personal use, study, research and other noncommercial use are free, including changing it and sharing what you changed, as long as the author credit and license notices stay exactly as they are and your version says it is based on vibepanel. Commercial use, whether inside a company, as a hosted service or bundled into a product, needs a separate license: write to jmr@jiangmuran.com.
Releases up to and including v1.20.1 were published under the MIT License, and copies obtained under it keep it.