Claude Agent Manager

See exactly what your Claude Code agents — and every subagent — are doing, live, without leaving VS Code.

100% local · read-only · zero config

When Claude Code spins up an orchestrator and a swarm of subagents, the terminal scrolls too fast to follow. Claude Agent Manager turns that firehose into a calm, live dashboard — the full agent tree, what each one is doing right now, what it changed, what it cost, and every conversation you've ever had — searchable in a keystroke.

Why you'll want it

• 👀 Never lose track of a subagent again. A live tree shows the orchestrator → every subagent, with status, current tool, and progress.
• 🧭 Understand a run at a glance instead of scrolling thousands of terminal lines.
• 🔎 Find any past conversation by what was said in it — across every project — and jump right back in.
• 💎 Beautiful, native, and safe. Pure VS Code styling, read-only, all on your machine. It never touches your running sessions.

✨ Features

🌳 Live agent & subagent tree

Your whole run, structured. The main/orchestrator agent with its subagents nested underneath — names and tasks pulled straight from Claude Code. Spinners for running, badges for done / error / waiting, token counts, and TodoWrite-based progress when it's real (never a fake bar).

⚡ "Right now" — the current tool call

For every running agent, see the exact tool in flight (Edit src/auth/session.ts, Bash npm test, Grep …) plus a live streaming activity log that updates as it works.

🔍 Search every conversation (full-text)

Hit the search icon and type a keyword. Claude Agent Manager scans the full text of every transcript on your machine with ripgrep — sub-second — and returns the matching conversations (subagent hits folded in), each with a highlighted snippet. Then Continue or Read it in one click.

▶️ Continue any session — your way

Pick up right where a session left off. The Continue button resumes it in Claude Code's editor view or a terminal (claude --resume) — your choice, from a dropdown that remembers your default.

📂 Session-wide changes

One combined, PR-style view of every file the whole run touched — across the orchestrator and all subagents — with +/− counts and one-click diffs of exactly what changed.

📖 Readable conversations (not raw JSON)

Open any agent's conversation as a clean, formatted Markdown view — prompts, replies, and tool calls — instead of squinting at raw .jsonl.

🖥️ Sidebar or full-screen editor

Live in the activity bar, or pop the whole dashboard out into a full-screen editor tab — both stay in sync.

🎛️ Three layouts, one click

Switch between Tree (hierarchy + inline detail), Cards (one live card per agent), and Split (master list + Activity / Files / Conversation tabs).

🔔 Status bar + finish notifications

A glanceable status-bar chip (✦ 2 running · 1.2M tok) even when the panel is closed, plus a toast when an agent finishes or errors.

🚀 Getting started

1. Install Claude Agent Manager from the Marketplace (or code --install-extension claude-agent-manager-x.y.z.vsix).
2. Click the Claude Agent Manager icon in the activity bar.
3. Run Claude Code as usual — ideally a task that spawns subagents — and watch it light up. ✨

🔎 Seeing nothing? First open is scoped to your current workspace. Click the scope toggle in the header (or run “Claude Agents: Toggle Scope”) to show all projects.

No setup, no API keys, no sign-in. It reads the transcripts Claude Code already writes under ~/.claude/projects.

⚙️ Settings

All under claudeAgentManager.*:

🛡️ Private & safe by design

• Read-only. It observes the transcripts Claude Code already writes — it never sends input to, pauses, or interferes with your agents.
• 100% local. No network calls, no telemetry, no accounts. Your conversations never leave your machine.
• Fast on huge runs. Multi-megabyte transcripts are tail-read incrementally, never loaded whole in the hot path.

~/.claude/projects/<project>/
  <sessionId>.jsonl            ← main session (the orchestrator agent)
  <sessionId>/subagents/
    agent-<id>.jsonl           ← one transcript per subagent
    agent-<id>.meta.json       ← { agentType, description }

🗺️ Roadmap

• 💵 Cost in $ (token usage × model pricing, per agent / session / day)
• 📊 Usage analytics (per-tool breakdown, context-window %, cache-hit rate)
• 🕓 Session history & trends

Ideas or bugs? Open an issue.

❓ FAQ

Does it work without any setup? Yes — if you use Claude Code, it just works.

Can it control or stop my agents? No, by design. Claude Code exposes no external control API, so this is a safe, read-only monitor. Continue launches a fresh resume via the CLI/extension; it never injects into a live run.

Does it send my data anywhere? Never. Everything stays on your machine.

“Continue → Claude Code (editor)” isn't resuming? It uses the official Claude Code extension; if that isn't installed, Continue falls back to a terminal automatically.

📄 License

MIT © Damilola Onadeinde