One task queue your parallel agent sessions claim work from.

Agents propose the work, you approve it once, and every session on the machine pulls from the same queue. It all runs locally: one binary, one SQLite file, no accounts.

npx -y @jaximus/tandem-mcp setup

Adds the MCP server to Claude Code, downloads the binary into ~/.tandem, starts it on 127.0.0.1:7891 and opens the board; restart Claude Code to pick it up. Other MCP clients run npx -y @jaximus/tandem-mcp directly.

How it works

When several agent sessions work on one repo, they usually share a TODO.md. A file can't track who is working on what. Tandem keeps tasks, claims and results in a queue that every session reads.

  1. An agent proposes an epic

    A named batch of tasks, each with a TDM-n ticket. Nothing runs yet.
  2. You approve it once

    On the board. That approves every task in it, and agents waiting on queue_wait start working.
  3. Sessions claim, atomically

    queue_next, then task_claim. A claim is one conditional UPDATE … WHERE state = 'approved'. If two sessions claim the same task, one gets it and the other is told who has it and takes the next one. A claim with no progress for 15 minutes expires.
  4. Each task completes with a commit

    The session reports a result and links the commit. The board updates live.

It all runs on your machine: one Go binary and one SQLite file in ~/.tandem. No accounts. Any MCP client works: Claude Code, Codex, Cursor, or a script of your own.