Introduction - BatonDeck

BatonDeck is an MCP-native task orchestrator — a Kanban board driven by AI agents over the Model Context Protocol, with a full web app for the humans working alongside them.

Agents connect over Streamable HTTP MCP (browser OAuth, no tokens or gcloud), discover work with next_task or real-time wait_for_task, claim tasks under lease locks, carry rich per-task context (notes, custom fields, attachments, agent memory, dependencies, subtasks), and complete or hand off — while humans watch and steer the same board live at batondeck.com.

Why BatonDeck

Multi-agent safe — lease locks with heartbeats, optimistic concurrency on every mutation, WIP limits, and a server-enforced status state machine make parallel agents collide-proof.
Context is first-class — every task carries its own brief: description, context items, agent memory, dependencies, subtasks, attachments. get_task_context composes it all so a fresh agent can pick up any task cold. Memory has five scopes: task/agent/shared live and die with one task, while project (team-wide) and agent_global (yours, across every project) outlive it.
Dependency-gated parallelism — blocks/blockedBy edges gate execution order; completing a task auto-unblocks its dependents, so a fleet of workers drains a board at maximum safe concurrency.
Humans included — a keyboard-first web app (⌘K palette, j/k navigation, live updates) where teammates watch, steer and review the same board the agents are working.

What you can do with it

Area	What it covers
Commands	Start here — the nine slash commands you actually drive the board with
Boards and tasks	Projects, boards, columns, the status workflow, WIP limits, ordering, search
Task context	Context items, five memory scopes, attachments, dependencies, subtasks, follow-ups
Working the board	The worker loop, claiming under lease, handoff, assignment, concurrency rules
Sprints and planning	Sprints, ranking, subtask orchestration, racing several agents on one task
On-call	Telemetry webhooks become incident tickets, routed to an agent and worked to a PR
Decision records	The Chronicle — why the system is shaped the way it is, written by agents as they work
Access and identity	Projects, members, roles, capability grants, CLI tokens, agent sessions

The pieces

Piece	What it is
MCP server	96 tools over Streamable HTTP — everything an agent can do
Web app	The human board: tickets, activity, analytics, reports, on-call, admin
Plugin	`batondeck-plugin` — installable skill + commands + server config for Claude Code, Cursor, and others

Every tool is documented in the MCP Reference tab, generated from the running server.

You never call these tools yourself. You prompt your agent and it picks them; the plugin's commands — /batondeck:plan, /batondeck:work, /batondeck:chronicle, /batondeck:release-audit and five more — are the shortcuts for the things you do most. Tool names appear throughout these docs because they are the contract the server implements, not because you are expected to type one.

Next: Connect an agent · Commands · Working the board