Claude Orchestrator

A kanban board for your Claude Code agents. Spawn a batch, watch them run live in a sidebar, get consolidated answers back in your chat — without leaving VS Code.

What it does

Most of the time you chat with a single Claude Code agent. Sometimes you want three: one to audit a repo, one to draft tests, one to research a library — at the same time. Or you want Opus, Sonnet and Haiku working the same prompt so you can compare answers.

Claude Orchestrator turns VS Code into the cockpit. Type a natural request in any Claude Code chat — "lanza 3 agentes paralelos: uno opus que audite src/, uno sonnet que esboce tests, uno haiku que resuma el README" — and the chat fans out the work, the sidebar dashboard tracks every agent live, and the consolidated answers come back to your conversation.

Why use it

• Compare models on the same task. Run Opus, Sonnet and Haiku in parallel and pick the winner.
• Parallelize research and audits. Three repos, three agents, one batch — finished in the time of the slowest one.
• Stay in control. Cancel any agent mid-flight from the dashboard, watch its live log in a detail panel, get a toast when each finishes.
• Keep your context clean. The orchestrator returns consolidated answers — not raw transcripts — so your chat's context window stays small.
• Pick the right model per task. Opus for architecture decisions, Sonnet for coding, Haiku for grep-and-summarize.

Install

1. Download claude-orchestrator-0.1.0.vsix from Releases.
2. Install it: code --install-extension claude-orchestrator-0.1.0.vsix (or right-click the .vsix in VS Code → Install Extension VSIX).
3. Reload VS Code.

That's it. The extension auto-registers with Claude Code on first activation — no token copy-pasting, no manual config.

Prerequisites: Claude Code CLI installed and signed in (claude login). VS Code 1.120+, Node 22+.

First spawn (60 seconds)

1. Open any terminal and start a Claude Code chat: claude.

2. Ask in natural language:

   Lanza 3 agentes paralelos en este repo: uno haiku que liste los archivos en src/, uno sonnet que me resuma el README, uno opus que sugiera 3 mejoras de arquitectura. Espera a que terminen y consolida.

3. Open the Claude Orchestrator sidebar in VS Code — UP NEXT → NOW PLAYING → RECENT fills as agents progress.

4. Click any card to open the detail panel with the live log.

5. The chat returns a consolidated answer from all 3 agents.

Screenshots and demo

Screenshots, demo GIF and a short walkthrough video are coming in v0.1.1.

What's inside

• 5 MCP tools exposed to any Claude Code chat: spawn_agents, wait_for_agents, list_agents, get_agent_log, cancel_agent.
• Kanban dashboard in the VS Code sidebar — NOW PLAYING / UP NEXT / RECENT, grouped by project and branch.
• Detail panel in an editor tab — per-agent streaming log with thinking, tool calls, tool results, usage and cost.
• Status bar with a live count of running agents (click to focus the dashboard).
• Per-task model selection — opus, sonnet, or haiku.
• Cancel from card buttons or via the cancel_agent MCP tool.
• Session resume — scans your ~/.claude/projects/ history and offers to resume past sessions in the chat or a new terminal.
• Safety caps — max agent runtime (default 33 min) and a stuck-detection signal that flags silent agents to your chat.
• Cost tracking in USD per agent — visible in cards, detail panel and the consolidated wait_for_agents result.

Configuration

Settings live under File → Preferences → Settings → Extensions → Claude Orchestrator. The defaults are sensible — touch these only if you're tuning.

Tips to get the orchestrator picked

Claude Code's tool selection sometimes prefers native Bash or Task over a custom MCP tool. Two simple things help a lot:

1. Name the tool in your prompt — "use spawn_agents from the claude-orchestrator MCP".

2. Add a directive to ~/.claude/CLAUDE.md so every chat picks the orchestrator by default:

   ## MCP server: claude-orchestrator
   
   When the user asks for parallel sub-agents, multi-model orchestration or
   batch delegation: prefer spawn_agents + wait_for_agents from the
   claude-orchestrator MCP server over Bash run_in_background or Task.
   The orchestrator gives a kanban dashboard, real cancel, model targeting
   and a long-poll fan-in.
   

You can also save a slash command at ~/.claude/commands/orchestrate.md and invoke /orchestrate <task> from any chat.

Advanced

claude mcp add-json --scope user claude-orchestrator \
  '{"type":"http","url":"http://127.0.0.1:39127/mcp","headers":{"Authorization":"Bearer <token>"}}'

git clone [email protected]:stevTresCloud/claude-orchestrator.git
cd claude-orchestrator
npm install
npm run compile      # bundle extension + webview
npm run watch        # rebuild on save
npm test             # vitest
npm run package      # produce .vsix

Caveats

A few things worth knowing before installing — full list in the CHANGELOG:

• The .vsix is ~75 MB because it bundles the official Anthropic CLI binary used at runtime. Distributed via GitHub Releases only (above the VS Marketplace 50 MB limit). v0.2 tracks slimming the bundle.
• First scan can take 60–90s if you have 150+ past Claude Code sessions; subsequent activations hit a cache and complete in ~25s.
• Single detail panel instance for v0.1; multi-instance with cap + LRU is on the v0.2 backlog.

License

MIT — see LICENSE.

Credits

Built on top of @anthropic-ai/claude-agent-sdk (official Anthropic SDK), with the multi-agent execution engine cherry-picked from damocles (MIT). MCP tools contract inspired by awslabs/cli-agent-orchestrator (Apache-2.0); dashboard UX inspired by manusa/ai-beacon (Apache-2.0); kanban schema informed by hoangsonww/Claude-Code-Agent-Monitor (MIT).

See THIRD-PARTY-NOTICES.md for the full list of bundled software.