Yaojia Wang c1c837c54f fix(relay-run): buffer early ws frames in wsToWebSocketLike
The message listener was attached only when the mux consumer called onMessage — which
happens inside attach(), AFTER the async mTLS-registry lookup. On loopback the agent's FIRST
heartbeat ping arrived in that gap and ws dropped it (no listener), so the agent's heartbeat
died on the first miss at 15s and the tunnel flapped forever. Attach the listener at
construction and buffer early frames (and an early close) until the consumer wires up.
2026-07-07 04:59:53 +02:00

Web Terminal

A self-hosted, browser-based terminal and Claude-Code session/project workbench. It exposes the host machine's local shell over WebSocket: open http://<host-ip>:3000 from any device on your LAN (phone, tablet, another laptop) and you get a live, interactive shell. The core use case is vibe coding — hand Claude Code a task, walk away, then reconnect from anywhere to check on it, approve a tool call from your phone's lock screen, or kick off the next one.

Sessions survive disconnects: the shell (and whatever's running in it) keeps going when you close the tab; reconnect and the scrollback replays. The server is a byte-shuttle — it ferries raw bytes between the shell and the browser and never parses terminal/ANSI semantics; xterm.js renders, node-pty provides the TTY.

⚠️ This hands a full shell to anyone who can reach the port. LAN-only, no authentication. Never port-forward or tunnel it to the public internet. See Security & deployment.


Features

Core terminal

  • WebSocket terminal — xterm.js in the browser, node-pty on the host. Keystrokes shuttle to the shell; output streams back. No terminal parsing on the server.
  • Sessions survive disconnects — the PTY lifecycle is decoupled from the WebSocket. Closing a tab detaches a client; the shell keeps running. Reconnect replays a ~2 MB scrollback ring buffer.
  • Auto-reconnect — exponential backoff (1s/2s/4s… capped at 30s), carrying the localStorage session id, so a flaky network or a phone waking from sleep just resumes.
  • Mobile touch key-bar — one-tap Esc / Esc·Esc / ⇧Tab / arrows / Enter / ^C / ^O / ^T / ^B / ^R / ^L / ^D / Tab / / — the high-frequency Claude Code keys a soft keyboard can't produce, sent as raw bytes (no soft-keyboard pop).
  • Multi-device session mirroring — many devices can attach to the same session at once. Output, exit and status broadcast to all; any device can type (shared control). PTY sizing is latest-writer-wins: whichever device you're actively using drives the size and stays full-screen.

Tabs & home

  • Multi-tab — each tab is an independent shell. Tabs auto-name to the current folder (double-click to rename), show a connection dot and an unread-output dot, and can be drag-reordered. + opens a new tab in the active tab's directory.
  • Home Sessions chooser — opening the app lands on a chooser, not a blank shell. It shows the host's running sessions as live preview thumbnails (read-only renders of each screen) so you can see what each one is doing, plus a dashed + New session tile. Pick one to open (full scrollback replay) or start fresh.
  • Per-session Open / Kill — manage sessions right from the chooser; no separate manage page.
  • Sessions ↔ Projects toggle — a segmented control on the home screen flips between the running-sessions view and the projects view.
  • ⌂ Home overlay — a Home button in the tab bar overlays the chooser on top of the current terminal so you can start another session/project without closing your tabs.

Claude Code cockpit

  • Live per-tab status — Claude Code hooks POST to the server (loopback side-channel); each tab badge shows working / waiting-for-approval / idle / stuck in real time. Install once with npm run setup-hooks.
  • Remote approve / reject — when Claude asks for tool permission, the request is held server-side and an Approve / Reject bar appears on every attached device — resolve it with a tap, no typing. Works across multiple devices (closing one mirror doesn't cancel the prompt for the others).
  • Plan-mode / permission-mode relay — start a session in a chosen --permission-mode (default / acceptEdits / plan / auto). When Claude exits plan mode, the approval bar becomes a three-way gate (approve+auto / approve+review / keep planning). The high-risk auto mode is gated behind ALLOW_AUTO_MODE.
  • tmux keepalive — run the shell inside tmux so sessions survive a server/host restart, not just a disconnect. Auto-detected, or forced with USE_TMUX.
  • Session history / resume — browse past Claude Code sessions (from ~/.claude/projects) and resume one.

Projects (v0.6)

  • Auto-discovered git repos — scans configurable roots for .git, showing each repo's branch and dirty state. Read-only, cached, with a depth-bounded BFS that skips node_modules/dotdirs/symlinks.
  • Per-project launchers — each card has brand-logo buttons: Claude and Codex open a new tab running that CLI in the repo; VS Code asks the host to open the editor on that path. Projects with an active Claude session highlight the Claude button (with a count).
  • Project detail page — branch + a git worktrees list, the project's active sessions (open/kill), a CLAUDE.md viewer with a Generate / Update (/init) button, a read-only git diff viewer, and a create-worktree action.

Walk-away workbench (v0.7)

  • Mobile Web Push + lock-screen triage — the host actively notifies your phone on needs-input (high priority, with Allow / Deny action buttons) and done (low priority). Approve or deny a held tool request straight from the lock screen without opening the app, secured by a per-decision capability token. Optional ntfy / Pushover bridge for setups without HTTPS Web Push.
  • Voice dictation — push-to-talk mic on the input bar (Web Speech API); release to send the transcript as input (optional auto-Enter). Hidden where unsupported; audio never touches the server.
  • Quick-reply chips + saved-prompt palette — tap-to-send chips (yes, continue, 1/2/3, Esc) plus a persistent palette of your own named snippets.
  • Activity timeline — a human-readable, timestamped stream of what happened while you were gone ("ran Bash · edited 3 files · waiting for approval · done"), built from discrete hook events in a bounded per-session ring.
  • Stuck / idle alert — if a session goes silent past a threshold (no output, not idle, not exited) it fires a one-shot "possibly stuck" alert through the same notification channel. Runs on the existing reaper tick (no extra timer).
  • statusLine telemetry → per-tab gauges — a statusLine script feeds back context-usage, cumulative cost, model, lines +/, PR state, and rate-limit (5h/7d) telemetry, rendered as per-tab gauges on tabs, thumbnails and project cards. Stale telemetry greys out.

UX

  • Themed UI (Amber dark theme + light / solarized, adjustable font size), scrollback search, an all-sessions dashboard, share-session QR/link (?join=<id>), connect-device QR, a keyboard-shortcut cheat-sheet, an installable PWA, clickable links, and a line-icon toolbar with hover tooltips.

Clients

The browser is the reference client; two native clients and a remote-access service are also in the repo (all consume the same wire protocol — the server stays a byte-shuttle).

  • iOS / iPad app (ios/, branch feat/ios-client) — a native SwiftUI + SwiftTerm pure remote client built for the walk-away loop: native terminal with scrollback replay, QR/manual pairing (Keychain), the remote Approve / Reject gate + three-way plan gate, APNs push with lock-screen Allow / Deny behind Face ID, deep links, Projects, multi-session switcher, timeline, diff, quick-reply, and an adaptive iPhone/iPad split layout. Looks like the desktop (amber-gold, dark-first). P0 needs zero server changes; P1 adds two declared additive touch-points. See ios/README.md. (Not yet merged.)
  • Desktop app (desktop/) — a Mac/Windows Electron shell that embeds this Node server + node-pty (all-in-one), for native notifications, tray, deep links and launch-at-login. Design in docs/DESKTOP_PLAN.md. (Built; packaged & launch- verified on Apple Silicon.)
  • Rendezvous relay (term-relay/ · agent/ · relay-e2e/ · relay-web/ · relay-auth/ · relay-contracts/) — an "ngrok-for-Claude-Code" with end-to-end encryption to reach a self-hosted host from anywhere, without a VPN or port-forwarding. Security-audited but pre-production / not deployed; see docs/PLAN_RELAY_INDEX.md and docs/DEPLOY_RELAY.md. For remote access today, use Tailscale (§ Security & deployment below).

Quick start

Prerequisites

  • Node.js ≥ 18 (developed on v24).
  • macOS / Linux. On macOS, node-pty compiles a native addon — install Xcode Command Line Tools (xcode-select --install) if npm install fails. After a major Node version bump, run npm rebuild.

Install & run

npm install        # installs deps (incl. node-pty native build); postinstall fixes the spawn-helper bit
npm run build:web  # bundle the frontend (public/main.ts → public/build/) with esbuild
npm start          # serve on 0.0.0.0:3000

Then find your LAN IP and open it from any device on the same network:

ipconfig getifaddr en0          # macOS
# open http://<that-ip>:3000

For frontend development, run npm run dev:web (esbuild --watch) alongside npm start.

npm run setup-hooks                 # adds the hooks + statusLine to ~/.claude/settings.json (backs it up)
# npm run setup-hooks -- --remove   # to uninstall

This wires Claude Code's hooks → live per-tab status, the statusLine gauges, and push notifications. The hooks are a no-op outside web-terminal (they only fire when $WEBTERM_* env vars are set in spawned shells), so they're safe to leave installed. Then run claude inside a tab.

USE_TMUX=1 npm start keeps sessions alive across a server restart.

Tests

npm test           # vitest, all modules (~470 tests, 80% coverage gate)
npm run typecheck  # tsc (backend + frontend)
npm run build      # compile backend to dist/

Configuration

All config is via environment variables (no hardcoding). Invalid values fail fast at startup.

Core

Var Default Purpose
PORT 3000 Listen port.
BIND_HOST 0.0.0.0 Listen address.
SHELL_PATH $SHELL or /bin/zsh Shell to spawn.
IDLE_TTL 86400 (s) Reclaim a detached session after this idle time (no new output since detach).
SCROLLBACK_BYTES 2097152 (2 MB) Per-session replay ring buffer size.
MAX_PAYLOAD_BYTES 1048576 (1 MB) Max single WS frame; oversized frames are rejected.
WS_PATH /term The only path accepted for WS upgrade.
MAX_SESSIONS 50 Cap on concurrent sessions (DoS guard).
MAX_MSGS_PER_SEC 2000 Per-connection WS frame-rate cap; over-limit frames are dropped (not disconnected).
USE_TMUX auto 1/0/auto — run the shell inside tmux (keepalive across restart); auto = on if tmux is on PATH.
ALLOWED_ORIGINS (derived) Extra allowed WS origins, comma-separated. The base list is derived from the host's NIC IPs + localhost — never from BIND_HOST.
PERM_TIMEOUT_MS 300000 (5 min) How long a held tool-permission request waits for a remote decision before falling back to Claude's own prompt. Must be > 0.
REAP_INTERVAL_MS 60000 Idle-reaper / stuck-sweep tick interval.
PREVIEW_BYTES 24576 (24 KB) Tail of scrollback served for live preview thumbnails.

Projects (v0.6)

Var Default Purpose
PROJECT_ROOTS $HOME Comma-separated absolute roots to scan for git repos (~/~/... expanded; relative paths rejected).
PROJECT_SCAN_DEPTH 4 How deep to descend looking for .git.
PROJECT_SCAN_TTL 10000 (ms) Cache TTL for repo discovery.
PROJECT_DIRTY_CHECK 1 (on) Run git status --porcelain to flag dirty repos.
EDITOR_CMD code Command the host runs for the project "VS Code" button.

Walk-away workbench (v0.7)

Var Default Purpose
VAPID_PUBLIC_KEY (unset → push off) Web Push public key, exposed to the service worker.
VAPID_PRIVATE_KEY (unset, secret) Web Push private key; required to enable push. Never logged.
VAPID_SUBJECT mailto:admin@localhost VAPID sub.
PUSH_STORE_PATH ~/.web-terminal-push-subs.json Persisted push-subscription store.
PUSH_MAX_SUBS 50 Max stored push subscriptions (DoS guard).
NOTIFY_DONE 1 (on) Send the low-priority DONE push on Stop/SessionEnd.
NOTIFY_DND 0 (off) Global do-not-disturb default.
DECISION_TOKEN_TTL_MS = PERM_TIMEOUT_MS Lifetime of a lock-screen decision capability token.
TIMELINE_MAX 200 Per-session activity-timeline event ring cap.
TIMELINE_ENABLED 1 (on) Capture/serve the activity timeline.
STUCK_TTL 600 (s) Silence window before a "possibly stuck" alert; 0 disables.
STUCK_ALERT 1 (on) Master switch for stuck alerts.
DIFF_TIMEOUT_MS 2000 Timeout for a single git diff.
DIFF_MAX_BYTES 2097152 (2 MB) Patch truncation cap.
DIFF_MAX_FILES 300 Max files returned before the diff is marked truncated.
STATUSLINE_TTL_MS 30000 After this with no update, per-tab telemetry greys out as stale.
WORKTREE_ENABLED 1 (on) Master switch for the create-worktree feature (the only write-to-disk action).
WORKTREE_ROOT (unset → <repo>-worktrees) Root that new worktrees must land inside.
WORKTREE_TIMEOUT_MS 10000 Timeout for git worktree add.
DEFAULT_PERMISSION_MODE default --permission-mode used when a session is started without an explicit choice (default/acceptEdits/plan/auto).
ALLOW_AUTO_MODE 0 (off) Whether the high-risk auto (bypass-permissions) mode is offered/honored.

Spawn-injected env (set by the server / read by hook scripts — not user config)

WEBTERM_SESSION, WEBTERM_HOOK_URL, WEBTERM_STATUSLINE_URL are injected into each spawned shell so the hooks/statusLine know which session they belong to and where to POST. The optional ntfy/Pushover bridge reads WEBTERM_NTFY_URL / WEBTERM_NTFY_TOPIC / WEBTERM_NTFY_TOKEN and WEBTERM_PUSHOVER_TOKEN / WEBTERM_PUSHOVER_USER — secrets stay in env, never written into settings.json or command argv.


Security & deployment

This is a no-auth, LAN-only tool by design — it hands a full shell to anyone who can reach the port. The defenses that matter:

  • WS Origin validation (cannot be skipped): the WebSocket handshake rejects any Origin not on the allow-list (HTTP 401). This blocks Cross-Site WebSocket Hijacking — a malicious page in your browser trying to connect to ws://<your-lan-ip>:3000. Only WS_PATH is accepted for upgrade.
  • CSRF / Origin guards on state-changing routes: every route with a side effect (DELETE /live-sessions, POST /open-in-editor, POST /push/subscribe, POST /hook/decision, POST /projects/worktree) requires an allowed Origin (403 otherwise). Read-only discovery routes don't.
  • Loopback-only hook ingest: the side-channel ingest endpoints (/hook, /hook/permission, /hook/status) only accept loopback peers — the Claude process always runs on the host.
  • Per-IP rate limits on push subscribe and lock-screen decision routes; a per-connection WS frame-rate cap.
  • Safe git exec + per-decision capability tokens: all git calls use execFile (no shell) with timeouts and path validation; the lock-screen Allow/Deny is authorized by a token bound to that session's current pending request (it expires on resolve/timeout).
  • No authentication yet. Treat the whole app as "shell access for anyone on the network." Never expose it to the public internet. ws:// is unencrypted, so on an untrusted network keystrokes (passwords, API keys) can be sniffed. The recommended deployment is Tailscale (WireGuard-encrypted), which also gives you wss:// — the frontend auto-selects wss on HTTPS. Web Push requires HTTPS (a secure context), so the push features need the Tailscale/TLS deploy; bare LAN-over-HTTP falls back to the ntfy/Pushover bridge.

Architecture

The server is a byte-shuttle, not a terminal: node-pty gives the shell a real pseudo-terminal, the server shuttles raw bytes over the WebSocket, and xterm.js in the browser interprets ANSI and renders. The server never parses terminal semantics — every "smart" feature (status, timeline, telemetry, diff, push) rides an out-of-band side-channel (loopback HTTP/JSON from Claude Code hooks, or git subprocesses), keeping the terminal stream a pure pipe.

The other central design point: PTY lifecycle ≠ WebSocket lifecycle. A WS close detaches a client (the PTY keeps running for other devices and for reconnect); only an idle timeout (or explicit kill / server shutdown) ends a session.

Tested with vitest (~470 tests, 80% coverage gate across backend + the logic-bearing frontend modules), plus real-PTY integration tests that auto-skip where posix_spawn is unavailable (sandboxes) and run everywhere else.

Design and rationale: docs/TECH_DOC.md (the why) and docs/ARCHITECTURE.md (the how). Feature PRDs: docs/FEATURE_PROJECT_MANAGER.md (v0.6) and docs/FEATURE_WALKAWAY_WORKBENCH.md (v0.7).

Description
No description provided
Readme 12 MiB
Languages
TypeScript 61.2%
Swift 36.7%
CSS 1.1%
JavaScript 0.8%
HTML 0.1%
Other 0.1%