dazhi666666

Agent Router — Development skill for Claude Code

Development community

本地多 Agent 协作调度器:主代理常驻主线程(Claude Code/ZCode/Devin/Codex/OpenCode/Antigravity 可指定),共享任务板 MCP、事件驱动非阻塞派发、自动收割合并.

How to install Agent Router

This entry records only its repository, not the path inside it, so there is no exact command to give. Open dazhi666666/agent-router and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Agent Router does

本地多 Agent 协作调度器:主代理常驻主线程(Claude Code/ZCode/Devin/Codex/OpenCode/Antigravity 可指定),共享任务板 MCP、事件驱动非阻塞派发、自动收割合并.

Alternatives in Development

  • Zeron — A native control plane for Claude Code, Codex, Cursor, Devin and other coding agents 1.4k ★
  • Oh My Agent — Portable multi-agent harness for .agents-based skills, workflows, and standards-aware agent teams across Antig 786 ★
  • Tulpa — 基于Windows本地QQ+微信的Harness/MCP:用Agent进行实时检索、跟进、管理、记忆与工作区 179 ★

README

Agent Router

English · 中文

**A main agent runs the main thread (Claude Code by default; pick zcode / devin / codex / opencode / antigravity via `--main`)**: it holds a long-lived session, breaks the overall goal into tasks and dispatches them to local subagents (ZCode / Devin / Codex / OpenCode / Antigravity / other Claude Code instances). Dispatch is non-blocking — the manager creates tasks and immediately moves on; subagent progress (completion / failure / messages) is pushed back into the main session in real time. All agents collaborate through a **shared local task board** and exchange messages via the **board mailbox**; the Router handles automatic dispatch, harvest & merge, and the final summary.

Everything stays on your machine — agent runtimes, sessions, task data and communication. Model inference calls each vendor's API; that is the only boundary.

Communication: shared task board + event-driven

Agents never talk to each other directly — everything goes through one **local MCP server** (stdio + a JSON file); the Router is the main agent's "hands and feet" (scheduler + event layer):

you ──goal──▶ main agent session (main thread, persistent, bidirectional stream-json)
                ├─ create_task → returns instantly, keeps working on its own stuff
                ├─ followup_task → a subagent continues with the original task context
                ├─ receives [event]/[mailbox] pushes → replies / spawns follow-ups / codes itself
                └─ goal achieved → calls finish_run to end the run
                      ▲ events injected live            │ dispatch / harvest (Node scheduler)
                ┌─────┴──────────────────────────────┴─────┐
                │        board-server.mjs (MCP/stdio)      │
                │   tasks + mailbox → data//board.json│
                └─────┬──────────┬───────────┬─────────────┘
              claude -p   agy(stream-json)  zcode -p   o