🎭 Claude Theater

Watch your Claude Code conversations and subagents work — a live office, in real time.

_{The demo office (claude-theater --demo): conversations and subagents at work, each agent walks in, sits, head-bobs while it types, then confetti on finish. A community visualizer for Claude Code — not affiliated with Anthropic.}

Try it now, no Claude Code session required — pipx run claude-theater --demo spins up the office above with synthetic agents.

Why? When you fan out a handful of subagents, the terminal becomes a wall of interleaved JSON. Claude Theater turns it into a glance — who's working, who's stuck, who just finished — without reading a single log line.

Safe by design: it runs entirely on your machine. Binds to 127.0.0.1, sends nothing anywhere (no telemetry), only reads the journals, and the whole thing is one auditable standard-library file with zero dependencies. (more)

Claude Theater reads the journals Claude Code writes for your conversations and the subagents they spawn, and renders them as a live office: every conversation is a room (titled by its subject), and inside it the conversation itself plus each subagent is a little character at a desk — avatar, name, the tool it's using right now, and a timer. Click any character to read its full task and result. When an agent finishes: confetti, a chime, and it quietly steps off the floor — and each room has its own show-finished toggle, so you control the history per conversation.

The UI is bilingual — English by default, Hebrew one click away (the choice is remembered, and the layout flips to RTL). Adding another language is a single edit to the I18N table in claude_theater.py; the server stays language-neutral.

Screenshots

Your real sessions — every conversation is a room; the conversation itself and each subagent is a character at a desk (avatar, name, the tool it's using now, a timer).

Bilingual — English ⇄ Hebrew (RTL). One click flips the whole office; the choice is remembered and the layout switches to RTL.

English	Hebrew (RTL)

Try it with no session — claude-theater --demo spins up the office with synthetic agents.

Quick start

Zero install — run it straight from PyPI with pipx:

pipx run claude-theater          # watches your real Claude Code sessions
pipx run claude-theater --demo   # a synthetic, populated office — no sessions needed

New here? Start with --demo to look around before pointing it at your own work.

Or install it:

pipx install claude-theater   # or: pip install claude-theater
claude-theater                # add --demo for the synthetic office

Or run from a clone (pure standard library, nothing to install):

python -m claude_theater

Then open http://localhost:7333. The server opens your browser for you — pass --no-browser to skip that (handy on a headless box or inside an editor panel). On Windows, from a clone, you can also just run start.cmd.

Requires Python 3.9+. The non-demo mode reads the journals a Claude Code install writes under ~/.claude/projects/; --demo needs nothing but Python.

VS Code extension

Prefer it inside your editor? The vscode-extension/ folder packages Claude Theater as a VS Code extension: it starts the server in the background (toggle it from a status-bar button) and opens the office in an interactive panel — no separate terminal. The server is bundled into the extension, so it works without a separate install, and it auto-starts when VS Code launches.

Install the prebuilt .vsix (easiest) — download claude-theater-<version>.vsix from the latest release, then in VS Code run Extensions: Install from VSIX…, or from a terminal:

code --install-extension claude-theater-<version>.vsix

Or build it from source (needs Node.js):

git clone https://github.com/asafabram-ship-it/claude-theater
cd claude-theater/vscode-extension
npx @vscode/vsce package        # produces claude-theater-<version>.vsix
code --install-extension claude-theater-*.vsix

Auto-start with Claude Code

The VS Code extension already starts the server on launch. If you use the CLI, the recommended one-time setup is a SessionStart hook in your ~/.claude/settings.json, so the office comes up by itself whenever you begin a Claude Code session. Each user adds it once. (Heads-up: this leaves an unauthenticated loopback endpoint running in the background — see Privacy.)

{
  "hooks": {
    "SessionStart": [
      { "hooks": [ { "type": "command",
        "command": "curl -sf http://127.0.0.1:7333/ -o /dev/null || (claude-theater --no-browser &)" } ] }
    ]
  }
}

{
  "hooks": {
    "SessionStart": [
      { "hooks": [ { "type": "command",
        "command": "powershell -NoProfile -Command \"if(-not(Get-NetTCPConnection -LocalPort 7333 -State Listen -ErrorAction SilentlyContinue)){Start-Process -WindowStyle Hidden -FilePath claude-theater -ArgumentList '--no-browser'}\"" } ] }
    ]
  }
}

Both check the port first, so they stay idempotent. (Requires claude-theater on your PATH, e.g. via pipx install claude-theater.)

How it works

• Polls the journals Claude Code already writes — both the session files (your conversations) and **/subagents/agent-*.jsonl (the subagents they spawn). Claude Theater only reads them; it never starts or controls agents.
• A single adapter, parse_agent_event(line) -> Event, is the only code that touches the raw journal format. Everything else consumes the stable Event, so a Claude Code format change is absorbed in one place.
• Degrades, never crashes: corrupt lines are skipped (and counted), unknown keys are ignored, and an agent missing fields still shows up instead of vanishing.
• A non-blocking version banner appears when a journal comes from a Claude Code version this build hasn't been tested against — output may be partial, but the office keeps running.

Privacy

Your journals contain real conversation content, so Claude Theater is built to keep them on your machine:

• Binds to 127.0.0.1 only — never reachable from the network, with a loopback Host allowlist that blocks DNS-rebinding.
• Never transmits anything anywhere. No telemetry, no remote calls.
• Read-only — it opens journals to read, never to write.
• A strict Content-Security-Policy and X-Content-Type-Options: nosniff, and cross-origin headers sent only to a vscode-webview:// origin, so an ordinary web page can't read your agents.
• The committed fixtures/ are 100% synthetic — no real prompts or results.

The local endpoint is unauthenticated: any process running as you on this machine can read it — the same trust boundary as the journal files themselves. The auto-start hook above leaves that endpoint up in the background.

Development

python -m unittest discover -s tests   # golden parser tests, zero dependencies

The tests run synthetic journal fixtures through the adapter and fail loudly if the format drifts. Contributing a journal sample from a new Claude Code build is the most valuable contribution — scrub every prompt and result first.

Compatibility

Tested against Claude Code 2.1.x. Newer builds will trigger the version banner; please open a format-drift report with a scrubbed sample so we can add a fixture and an adapter case.

License

MIT © 2026 Asaf Abramzon. "Claude" is a trademark of Anthropic; this is an independent, community project for Claude Code.