AgentDock

Stop managing AI coding agents like terminal tabs.

Mission control for parallel Claude Code and Cursor Agent sessions. Run a fleet across repos from one local-first cockpit — desktop or phone. See who is working, who needs you, review the diff, and jump in.

• Parallel Claude Code and Cursor sessions
• Isolated git worktrees
• Live working / waiting / done status
• Plans and diffs before you merge
• The same UI on your phone
• Local-first — your machine, no cloud

Install

curl -fsSL https://raw.githubusercontent.com/vishalnarkhede/agentdock/main/install.sh | bash
agentdock web

Or clone it yourself:

git clone https://github.com/vishalnarkhede/agentdock.git
cd agentdock
./setup.sh
agentdock web

./setup.sh installs Bun and tmux if they are missing, then prints a checklist. Claude or Cursor is only required when you start an agent — the UI opens without them.

  ✓ git
  ✓ bun
  ✓ tmux
  ! claude or agent (Cursor) — needed to start a session
    https://docs.anthropic.com/en/docs/claude-code
    https://docs.cursor.com/cli/agent

Re-run the checklist anytime with agentdock doctor.

First launch asks where your repos live and which ones to add. Password, Linear, Slack, and ngrok are optional and live in Settings.

Opens http://localhost:5173 (API on port 4800).

What changes

Before

Terminal 1: Claude fixing auth
Terminal 2: Claude doing tests
Terminal 3: another repo
Terminal 4: forgotten agent waiting for permission
Terminal 5: git diff
Phone: useless

After

One dashboard
  3 working
  1 needs input
  2 ready for review

Three ways people use it

Run a team of tasks. Start several Claude or Cursor sessions across one repo or many. Isolated worktrees keep branches from colliding. The queue shows working / waiting / done — Claude Code hooks where they exist, terminal fallback for Cursor.

Review before you merge. Open the plan the agent wrote, read the live diff, comment, and send a follow-up without leaving the session. Restore a stopped session and the conversation comes back.

Walk away. The same UI is built for a phone. Start four jobs on a laptop, close it, answer the one that needs input from your pocket. On the LAN use https://<your-ip>:5173; away from home, toggle ngrok in the header.

Why this exists

I got tired of Cursor eating RAM, of not being a terminal person, and of features that span four repos. Sometimes I just need things to keep moving while walking the baby at 2am. If that sounds familiar, this is for you — say hello on X.

AgentDock is under active development. Bug reports and feedback are welcome.

CLI

agentdock start my-repo                # launch an agent
agentdock start -r repo1 repo2         # one session, several repos
agentdock start my-repo --isolated     # isolated git worktree
agentdock doctor                       # what is missing
agentdock repos                        # configured repos
agentdock list                         # active sessions
agentdock stop --all                   # kill them

Using from your phone

1. agentdock web on the Mac or Linux box (it binds on the LAN).
2. Your IP: ipconfig getifaddr en0 (macOS) or hostname -I (Linux).
3. On the phone, open https://<your-ip>:5173. The self-signed cert warning is expected on the local network — Advanced → Proceed.
4. Add to the home screen if you want it full-screen.

Away from home, turn on ngrok in the header and open the public HTTPS URL.

Configuration

Everything lives in ~/.config/agentdock/. No database.

Repos are added on first launch or later under Settings > Repositories (scan a folder, add by path, remove). Or write repos.json yourself:

[
  { "alias": "backend", "folder": "my-backend", "remote": "org/my-backend" },
  { "alias": "frontend", "folder": "my-frontend", "remote": "org/my-frontend" }
]

Default base directory is ~/projects. Override with AGENTDOCK_BASE_PATH.

File	Purpose
~/.config/agentdock/auth-password	Optional password for the web UI

MCP servers (for example Linear) are added under Settings > MCP Servers and synced into agent configs.

Cortex is a good companion if you want Claude sessions to remember things across restarts.

How it works

See ARCHITECTURE.md for internals.

1. Create a session — pick repos, optionally isolate with a worktree.
2. The agent starts in tmux with tools for files, git, and gh.
3. Status comes from Claude Code hooks (PreToolUse, Stop, Notification, …) written to /tmp/agentdock-status/.
4. The browser attaches through a PTY and streams the real terminal over WebSocket.
5. Plans land in ~/.config/agentdock/plans/. The Changes tab is a live git diff.
6. Restore finds the last Claude conversation and runs claude --resume <uuid>.
7. Stop kills tmux and removes worktrees you asked it to create.

Keyboard shortcuts

Shortcut	Action
Ctrl+Shift+[ / Ctrl+Shift+]	Previous / next session (MRU)
Cmd+F	Search inside a file (Files tab)
Esc	Collapse the plan / changes / files panel

Also built in

Switch Claude ↔ Cursor mid-conversation, session templates, pinning and grouping, custom session properties, Slack-to-fix, browser notifications, optional password. To plug in another CLI agent, see AGENTS.md.

Security

Local-first. The server binds to your machine. State is files, not a hosted database. Optional hashed password in ~/.config/agentdock/auth-password. Each session is its own tmux session with configurable tools.

Report vulnerabilities via SECURITY.md.

Platform

Platform	Status
macOS	Supported
Linux (Ubuntu/Debian)	Supported
Windows	WSL only (tmux)

FAQ

Why tmux? Persistent sessions and a real terminal without writing a process supervisor.

Why Bun? Fast start, TypeScript, native WebSockets.

Why no database? tmux holds the live session; a few JSON files hold config.

Remote access? Set a password and use the LAN, ngrok, SSH, or a reverse proxy.

Status? Claude hooks write /tmp/agentdock-status/. Cursor falls back to matching the terminal.

Restore? Latest conversation UUID from ~/.claude/projects/ is passed to claude --resume.

Contributing

See CONTRIBUTING.md.

License

MIT — LICENSE.