Real-time X-ray for Claude Code agent swarms

What it does The project wires Claude Code’s 12 lifecycle hooks—like PreToolUse, SubagentStart, and SessionEnd—into Python scripts that HTTP POST events to a Bun server. A Vue client then displays the firehose in real time: tool calls, subagent handoffs, permission requests, and session lifecycles across multiple concurrent agents. It is essentially glue between Claude Code’s internal telemetry and a web dashboard, with SQLite acting as the buffer.

The interesting bit The hooks do more than just log; some actively gatekeep. pre_tool_use.py blocks dangerous commands, user_prompt_submit.py can reject prompts with a JSON {"decision": "block"} response, and stop.py carries a stop_hook_active guard to prevent infinite loops. That turns the observability pipeline into a lightweight policy layer, not just a passive stream.

Key highlights

• Captures all 12 Claude Code hook event types, including MCP tool detection and subagent transcript paths
• Bun server with SQLite (WAL mode) and WebSocket /stream for live updates
• Vue 3 client with event timeline, multi-select filters, live pulse chart, and chat transcript viewer
• Includes pre-built agent team definitions (builder.md, validator.md) and a team-planning slash command
• Hook scripts forward event-specific fields (tool_name, agent_id, notification_type) as top-level properties for easier querying

Caveats

• Setup requires manually copying the .claude directory and editing settings.json for each project you want to monitor
• The README lists optional API keys for OpenAI, ElevenLabs, and Firecrawl, but leaves it unclear which features degrade if they are absent

Verdict Worth a look if you are already running Claude Code with multiple agents or subagents and need visibility into the mess. Skip it if you just want a managed SaaS observability platform; this is a self-hosted wiring project.