dispatch

/dispatch

Manage agent dispatches via the AgentDispatch MCP server.

Quick Reference

Tool	Purpose
mcp__dispatch__run	Dispatch headless agent (blocks until done)
mcp__dispatch__interactive	Open interactive terminal session (non-blocking)
mcp__dispatch__status	Check active dispatches
mcp__dispatch__result	Get completed result
mcp__dispatch__cancel	Cancel running dispatch
mcp__dispatch__config	Show available models/agents/providers

Dispatching an Agent

mcp__dispatch__run(
  query="Fix auth bug",           # Short label (max 20 chars)
  prompt="Fix the JWT validation  # Full task description
    bug in auth.py line 42...",
  agent="implementer",            # Optional — general if omitted
  model="sonnet"                  # Optional — config default if omitted
)

Model Selection

• Anthropic (via Claude CLI): "sonnet", "opus", "haiku"
• Ollama (via OpenCode): "ollama/qwen3:30b", "ollama/qwen3:235b"
• OpenAI (via OpenCode): "openai/gpt-4.1"
• Groq (via OpenCode): "groq/llama-4-scout"
• Empty = uses default_model from dispatch.json

Built-in Agents

• general — Any task, full tool access
• explore — Read-only search and analysis
• reviewer — Code review, read-only
• implementer — Code implementation, full tool access
• raw — No agent definition, prompt goes directly
• Custom: pass file path like "./my-agents/security.md"

Parallel Dispatch

Call mcp__dispatch__run multiple times in one message — they execute concurrently:

# These run in parallel:
mcp__dispatch__run(query="Fix auth", prompt="...", agent="implementer", model="sonnet")
mcp__dispatch__run(query="Add tests", prompt="...", agent="implementer", model="sonnet")
mcp__dispatch__run(query="Review API", prompt="...", agent="reviewer", model="haiku")

Interactive Sessions

For tasks needing human input, use mcp__dispatch__interactive:

• Opens new terminal pane (Windows Terminal split / tmux split)
• Falls back to new window if not in WT or tmux
• User interacts directly with the agent
• Check completion via mcp__dispatch__status