Let Your Claude Code Agents Talk to Each Other: Introducing agent-dispatch 🤖↔️🤖

Let Your Claude Code Agents Talk to Each Other: Introducing agent-dispatch 🤖↔️🤖

TL;DR: An MCP server that lets Claude Code agents delegate tasks to specialized agents in other project directories. No more context-copying or permission sprawl. GitHub | PyPI

The Problem: Agents Live in Silos 🏝️

Working with Claude Code is fantastic until you need to cross project boundaries. If I'm debugging backend/ but need to check infra/ logs or query a staging DB, I'm forced to:

1. Copy-paste context → fragile, loses project-specific config.
2. Give one agent access to everything → security risk, prompt bloat.
3. Switch terminals manually → breaks flow, defeats the AI assistant.

I wanted specialized agents that collaborate without sharing credentials, configs, or context pollution. So I built agent-dispatch.

What Is It? 🚀

agent-dispatch is an MCP server that turns each project directory into an isolated, callable agent. When you dispatch a task, it spawns a fresh claude -p session in that directory, inherits its local CLAUDE.md, .mcp.json, and tools, executes the task, and returns a structured result.

Think of it as microservices for AI workflows: composable, cached, and sandboxed.

Quick Start (30s)

pip install agent-dispatch
agent-dispatch init
agent-dispatch add infra ~/projects/infra
agent-dispatch test infra "List running Docker containers"

That’s it. Now every Claude session can dispatch tasks to your infra agent.

How It Works 🔧

Your Claude session (e.g., backend/)
  │
  ├─ dispatch("infra", "find scheduler errors", caller="backend")
  │
  ▼
agent-dispatch MCP server
  ├─ ✅ Cache check → hit? return instantly
  ├─ 🔒 Safety check (depth, budget, concurrency)
  └─ 🚀 subprocess.run("claude -p ...", cwd=~/projects/infra/)
       │
       ▼
     Fresh Claude session in infra/
       ├─ Loads infra's CLAUDE.md, .mcp.json, tools
       ├─ Receives structured prompt: goal + caller + context + task
       └─ Executes → returns result → cached for next time

The Toolkit 🧰

Example call:

{
  "agent": "db",
  "task": "Are all migrations applied?",
  "caller": "backend",
  "goal": "Debug startup failure"
}

Safety & Cost Control 🛡️

I didn't want a demo toy—I wanted production-ready tooling. Built-in safeguards:

• Recursion protection: max_dispatch_depth (default: 3)
• Cost limits: max_budget_usd per agent or globally
• Concurrency caps: max_concurrency limits parallel claude -p processes
• Timeouts: kills stuck sessions (default: 300s)
• Caching: identical (agent, task, context) requests return instantly. Only successes cached.
• Hot-reload config: add/remove agents via ~/.config/agent-dispatch/agents.yaml without restarting.

When to Use (and When Not To) ⚖️

✅ Do dispatch when:

• The task needs tools, files, or context from another project
• You want to leverage project-specific MCP servers or CLAUDE.md
• You need parallel answers from multiple codebases

❌ Don't dispatch when:

• The task is simple and local (just use your current agent)
• You need sub-second latency
• You're in a tight loop (use dispatch_parallel instead)

Try It & Let Me Know! 💬

🔗 GitHub

📦 PyPI

pip install agent-dispatch
agent-dispatch init
agent-dispatch add my-agent ~/path/to/project

I'd love your feedback:

• What agent collaboration patterns would you use this for?
• What safety features are missing?
• Would you prefer a different abstraction (e.g., HTTP API vs MCP)?

Drop a comment below or open an issue. This is very much a "build in public" project. 🙏