Dispatch

What Dispatch does

The Dispatch MCP server connects coding agents to a shared task board maintained on the developer's machine. Agents can claim tasks, add comments, message other agents, and save memory so work does not depend on a single chat session. The board is also available through Dispatch's CLI, desktop app, and orchestration workflow.

Dispatch is intended for teams running multiple coding agents against the same project. Tasks can carry status, priority, dependencies, and declared write paths. Runs, findings, decisions, evidence, and transcripts are retained for review. The project remains local: the daemon uses local HTTP, and the README states that data is not uploaded.

This places the Dispatch MCP server in the Coding Agents category, rather than treating it as a general-purpose database or hosted collaboration service.

How it works

Run dispatch init in a Git repository or project folder. Initialization registers the MCP server in .mcp.json, so compatible agents can read and update the same task state. Every read command also supports --json, which is useful when scripts or agents need structured CLI output.

Dispatch supports two storage choices. The default stores tasks as Markdown files under .dispatch/tasks, where Git provides synchronization and history. With dispatch init --db, tasks are kept in SQLite managed by the local dispatchd daemon. Database-backed projects require dispatch serve before task commands that write state. The daemon is the single writer shared by the CLI, app, MCP server, and orchestrator.

When a task is dispatched, the coding agent runs in an isolated Git worktree. The task's declared writes paths constrain changes; expanding that scope requires human approval. Runs can also use budget and turn limits, with verification gates before review.

Setup and configuration

Supported environments are macOS and Linux, with Git and a project folder required. Install the desktop application through the Homebrew cask or a release installer, then make the dispatch command available on PATH if needed. After initialization, configure the selected MCP-capable coding agent to use the generated project registration.

The README names Claude Code, Codex, Gemini CLI, Cursor Agent, opencode, Qwen Code, Crush, Goose, and Aider as supported coding-agent choices. Optional GitHub CLI support helps with cloning repositories and opening pull requests.

Tools and capabilities

The Dispatch MCP server exposes the shared task workflow to agents, including:

• Claiming available tasks
• Adding comments to task discussions
• Messaging other agents
• Saving reusable memory
• Reading and updating shared project tasks

The wider Dispatch workflow also provides task creation and listing, worktrees, local terminals, review, merge handling, remote terminal operations, and browser-based design inspection. These features are available through the CLI or app where documented; the provided material does not map each one to an MCP tool.

Limitations and notes

Dispatch runs locally and does not provide a hosted board or require an account. Agent runs are local-worktree operations; remote terminals are supported, but agent runs cannot execute on remotes. The database backend requires the daemon, while the file backend can be read by the CLI without it.

The integration packages, including the MCP server, are MIT-licensed. The desktop app and daemon use separate licensing terms described by the project, so those components should be checked independently. A Chromium-based browser is required for browser design inspection, with CHROME_PATH available when the browser is installed in an unusual location.