Agent Radar Claude Code mod: a Claude Code subagent monitor above the input

Agent Radar is a community mod by Hamza Zafar (hamzafer), part of his claude-code-mods collection. It is a Claude Code subagent monitor: when Claude hands work to a subagent, the band above the prompt gets one live line for it, so you can see what each one is doing without opening anything.

A line looks like ● Read README and all src/ files 4s · 2 tools · searching src/**/*. It has the agent's description, the elapsed time, how many tool calls it has made, and a plain label for the latest call. When the agent finishes, the dot turns into a green check for 30 seconds, and a toast such as ✓ sweep runner done (9s · 4 tools) appears. A failed or aborted agent gets a red cross and the reason.

/radar opens a pane titled "Agent Radar". It lists up to nine agents with a hotkey each, shows how many are running and how many are finished, and lets you open one to see its last six messages. The pane has a "Clear finished" button. Agents that started before the mod loaded are picked up from the engine's own agent list when you run /radar. The mod does not show token counts or cost.

The recording shows Claude Code 2.1.291 with Haiku in a small demo project. We ask Claude to use an Explore subagent named scan src to read README.md and each file in src/. Claude starts it as a background agent, and a yellow line appears above the prompt at once: ● scan src 4s · 2 tools · searching src/**/*. The time and the tool count go up every second, and the label changes to reading src/login.js as the subagent moves from one file to the next.

The install block above adds the author's marketplace and installs the plugin from it. The band appears the next time a subagent starts. After installing into a session that is already open, run /reload-plugins.

The mod keeps the radar state in three host-held $.state atoms (agents, selected, now), so it survives a hot reload. Four hooks feed it. agent.spawn adds an entry, tool.call counts tool calls that carry an agentId, and turn.complete marks the agent done or failed and raises the toast. A one-second clock started on session.start redraws the elapsed times while any line is shown. The tool-call hook, from hooks/register.tsx:

on('tool.call', async ($, e, next) => {
  if (e.agentId) {
    const id = e.agentId
    const line = describe(e as unknown as Record<string, unknown>)
    await update($, agents, list => list.map(a => (a.id === id ? { ...a, tools: a.tools + 1, last: line } : a)))
  }
  return next(e)
})

It updates the counter before it lets the call run, and then returns the result of next(e) untouched. Calls from the main agent, which have no agentId, are ignored. describe turns a call into a short label: reading and editing plus the last two path segments, searching plus the first 30 characters of a pattern, running plus a Bash description or the first 50 characters of the command, and browser: click for Playwright tools. The band is drawn on the AbovePrompt render site after the content that other mods and Claude Code put there, and it steps aside while a survey is open.

There are no options in /config. To change the look, edit the constants at the top of register.tsx: SHOW_DONE_MS (how long a finished line stays, 30 seconds) and BAND_ROWS (how many lines the band shows, 4). The agent list is capped at 30 entries in the agent.spawn hook, and colors and glyphs are set in the icon and color functions.

We read every file the plugin loads before running anything: .claude-plugin/plugin.json, hooks/hooks.json, hooks/register.tsx, types/index.d.ts and the test file. There is no package.json, install script, binary or bundled code, and we found no hidden or zero-width characters. The repository has an MIT LICENSE.

Changes the UI covers the band, the pane and the toasts. Adds a command covers /radar. The mod uses no network, shell commands, file writes or environment variables, and it stores nothing outside Claude Code's host state, which is held in memory. It starts no background service: the one-second clock belongs to the session and runs only while an agent line is shown. It adds nothing to prompts or tool results and never calls the model.

The notable points are these. The mod hooks every tool.call and agent.spawn with no matcher, so it sees subagent tool names and inputs, such as a Bash command or a file path, and keeps a short label of the latest one in memory. It does not save or send them. In the pane, opening an agent reads its recent messages through $.session.messages and shows them locally, cut to 400 characters each. Last, claude plugin validate reported no errors and no warnings, but noted that the agent.spawn and tool.call hooks have no .catch. Both hooks pass the event on with next(e), so the practical risk is a thrown error from a state update, not blocking. The author's 2 tests passed on Claude Code 2.1.291.

Checked with Claude Code 2.1.291 on macOS (validate and tests). If the band does not show:

• A subagent has to start in the current session; the band only lists agents it saw spawn. Run /radar to pull in agents that started earlier.
• A survey is open above the prompt; the band returns once it closes.
• Another plugin already owns /radar. Claude Code refuses the name and the mod carries on without the command.
• Installed into an open session: run /reload-plugins.

The "Agent view" footer of Claude Code is separate from this mod and shows its own list of agents below the prompt.