Developer

How simple10 Built Agents Observe to Watch Claude Code Agents Live

Built by simple10 · · Source: GitHub

Quick answer

simple10 built Agents Observe, an open source dashboard that shows what Claude Code and Codex agents are doing as it happens. It streams tool calls, shows which subagent came from which parent, and lets you search and replay past sessions.

Result: The README states no results. The repo page shows 735 GitHub stars and 86 forks.

What they built

simple10 built Agents Observe, an MIT-licensed dashboard for watching Claude Code and Codex agents in real time. The README argues that when Claude Code runs on its own with subagents, the terminal shows only a fraction of what is going on. It treats tool calls as the ground truth, because the assistant's text is only a summary of what happened.

According to the README, the dashboard lets you:

  • Watch tool calls stream in as they happen, first the call and then the result
  • See the full agent hierarchy, including which parent spawned which subagent
  • Filter by agent or tool type, or search across all events
  • Expand any event to see its full payload, command and result
  • Browse past sessions, which get readable names like twinkly-hugging-dragon
  • Check token usage and cost breakdowns in session stats

How it works

The README describes a pipeline with four parts. Claude Code hooks fire on events like tool calls, prompts, stops and subagent starts. A small Node script reads each event and posts it to a local server. The server, built with Hono, saves everything in SQLite and forwards events over WebSocket to a React dashboard.

The server stores agent metadata and parentage, and tracks whether a session is active or stopped. It does not track agent status. The dashboard works that out itself from the event stream. Code that is specific to one kind of agent lives in the hook scripts, so the server stays the same no matter which agent sends the events.

Setup is two commands for the recommended plugin route: add the simple10/agents-observe marketplace in Claude Code, then install the agents-observe plugin. After that you run claude and open localhost:4981. The plugin also adds an /observe skill with subcommands such as view, stats, status and logs. The README lists Docker, Node.js and Bash as prerequisites, and there is a standalone route for development that uses the just task runner.

Codex support is checked off on the roadmap. OpenClaw and pi-code support are listed as still to come. simple10 credits an earlier project, Multi-Agent Observability System, as the inspiration.

The result

The README gives no numbers or benchmarks. The repo page shows 735 stars, 86 forks and 892 commits.

How to build something similar

  • Start with hooks. Claude Code can run a script on events like tool calls and subagent stops. Have that script send each raw event to a small local server.
  • Store raw events first. Save every event untouched in something like SQLite. You can work out agent trees, timelines and costs from the log later.
  • Push updates to the browser. Load history with a normal request, then use a WebSocket for new events so the page updates as the agents work.
  • Add filters early. Multi-agent runs produce a lot of events. Filtering by agent and tool type makes the stream readable.

The original post

simple10 · View on GitHub

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases