herdr-watcher
A small web dashboard for herdr agents, built for your phone. Project page: lab486.io/herdr-watcher.
- Every workspace is a tab. All comes first and shows every agent.
- Each agent shows herdr's status circle: working, needs you, done, or idle.
- When Claude asks for permission or asks a question, the options become buttons.
- Decisions walks you through every waiting question, one at a time.
- Open an agent to read the conversation, see its terminal, or send it a message and keys.
The server listens on 127.0.0.1 only. How you reach it from your phone is your choice: Tailscale, an SSH tunnel, a reverse proxy, remote desktop (RDP), or anything else.
Install
Requires herdr 0.9+ and Node 20+ on macOS, Linux or Windows.
herdr plugin install lab486/herdr-watcher
The server starts with herdr. Run the Watcher: open dashboard action, or go to http://127.0.0.1:7483.
To develop from a checkout:
npm run build
herdr plugin link "$PWD"
herdr plugin action invoke herdr-watcher.restart
Reach it from your phone
Put any tunnel or proxy in front of port 7483. For example, with Tailscale:
tailscale serve --bg 7483
The server refuses requests whose Host it doesn't know, which stops DNS-rebinding attacks. Add your tunnel's hostname to config.json in the plugin config directory. Run herdr plugin config-dir herdr-watcher to find the directory.
{ "port": 7483, "allowedHosts": ["my-mac.tail1234.ts.net"] }
Then run Watcher: restart server.
Security. Anyone who can load this page can type into your terminals. The watcher has no login of its own. Only expose it through something that already authenticates you, such as a tailnet or an SSH tunnel.
How it works
server/ is plain Node with no runtime dependencies. It polls herdr's socket (session.snapshot) once a second while a browser is connected, and it streams state over Server-Sent Events.- Herdr knows that an agent is blocked, but it does not know the question. For blocked Claude panes,
server/prompt-claude.js reads the visible screen and parses the dialog into options.
- An answer includes a fingerprint of the question. The server reads the screen again before it sends keys, so a stale tap can't answer a different question.
- The chat view reads Claude's session transcript from
~/.claude/projects/*/<session-id>.jsonl.
web/ is React, Tailwind, shadcn/ui, AI Elements and xterm.js.
Prompt parsing supports Claude Code only. Other agents still show their status, terminal and composer.
Develop
npm test # parser, transcript, and request-guard tests
HERDR_WATCHER_PORT=7483 npm start # server (serves web/dist)
npm run dev # Vite dev server, proxies /api to :7483
Parser fixtures in test/fixtures/ are real screen captures from herdr pane read --source visible --format text. When Claude changes a dialog, capture it again and add a test.
Built by Lab486. Want help with coding-agent workflows? Email [email protected].
