Skip to content
Dispatch

Introduction

Applies to: Dispatch v1.0.0, last verified 2026-04-05

Dispatch is a real-time monitoring and orchestration dashboard for AI coding agents. It supports four providers (Claude Code, OpenAI Codex, GitHub Copilot, and Google Gemini) through a single interface. When you run multiple agents in parallel across separate terminals (or spawn them directly from the dashboard), Dispatch shows you what every agent is doing so you don’t have to flip between windows or grep through files yourself.

What Dispatch does

Section titled “What Dispatch does”
  • Watches the .tasks/ directory that agents use to coordinate work
  • Tracks sessions from all four providers: Claude Code (real-time file watcher), Codex, Copilot, and Gemini (30-second poller)
  • Shows live status for every agent: what phase they are on, whether they are thinking, running a tool, or idle
  • Exposes an MCP server with six tools that agents call natively to query state, update phases, and emit activity events
  • Spawns CLI agents directly from the dashboard via the Run button on epic phase cards (Claude, Codex, Gemini)
  • Streams completed phase summaries into an activity feed as agents finish work
  • Provides a single browser tab where you can monitor any number of agents across any number of projects
  • Lets you filter the entire dashboard by provider using URL-param-driven pills (?providers=claude,codex)

What Dispatch does not do

Section titled “What Dispatch does not do”
  • Modify any files that agents write (epic data, execution logs, session files remain agent-owned)
  • Require agents to communicate with it directly. Agents write files; Dispatch watches those files. The MCP tools are optional.
  • Act as a message broker between agents. Agents coordinate through the shared .tasks/ filesystem, not through Dispatch.

How it works (overview)

Section titled “How it works (overview)”

Dispatch has five layers: agents write files to .tasks/, the Express server watches those files and builds state, a multi-provider session tracker reads logs from all four providers, an MCP server exposes tools that agents call natively, and a React dashboard displays everything in real time over SSE. An optional executor layer lets you spawn agents from the dashboard.

See How It Works for the full conceptual walkthrough: the file protocol, heartbeat thresholds, phase lifecycle, and data flow end to end.

Recovering work that agents left unfinished

Section titled “Recovering work that agents left unfinished”

When an agent runs out of context, crashes, or gets interrupted, its phase stays IN_PROGRESS or TODO indefinitely. The epic stalls. The Missions view still shows it as in progress. You may not notice for days.

The Orders view exists to surface this immediately. It flattens every phase from every epic into a single list, sorted so phases with no live agent — the abandoned ones — appear at the top. Click a phase to read the execution log the previous agent left behind, then hand the phase to a fresh agent. Work continues from exactly where the last agent stopped.

This is the feature that makes Dispatch useful beyond a basic liveness dashboard. See Epics and Tasks for the full walkthrough.

When Dispatch is useful

Section titled “When Dispatch is useful”

Dispatch is most useful when you have multiple agents running in parallel on different phases of the same work. With one agent, it mostly shows you what you already see in the terminal. With three or more agents splitting phases across a plan (especially a mixed fleet of Claude, Codex, and Gemini agents), the dashboard becomes the only practical way to track who is doing what.

Requirements

Section titled “Requirements”
  • Node.js >= 20
  • A project using the .tasks/ coordination convention (or any .tasks/ directory in the format Dispatch expects)
  • At least one supported CLI agent installed: claude, codex, gemini, or gh (Copilot). The dashboard works with zero agents running, but session tracking requires at least one provider.